Today, convolutional neural network (CNN) pruning techniques often rely on manually crafted importance criteria and pruning structures. Due to their heuristic nature, these methods may lack generality, and their performance is not guaranteed. In this paper, we propose a theoretical framework to address this challenge by leveraging the concept of $\gamma$γ-weak submodularity, based on a new efficient importance function. By deriving an upper bound on the absolute error in the layer subsequent to the pruned layer, we formulate the importance function as a $\gamma$γ-weakly submodular function. This formulation enables the development of an easy-to-implement, low-complexity, and data-free oblivious algorithm for selecting filters to be removed from a convolutional layer. Extensive experiments show that our method outperforms state-of-the-art benchmark networks across various datasets, with a computational cost comparable to the simplest pruning techniques, such as $l_{2}$l2-norm pruning. Notably, the proposed method achieves an accuracy of 76.52%, compared to 75.15% for the overall best baseline, with a 25.5% reduction in network parameters. According to our proposed resource-efficiency metric for pruning methods, the ACLI approach demonstrates orders-of-magnitude higher efficiency than the other baselines, while maintaining competitive accuracy.

RGB-Thermal (RGB-T) tracking enhances visual tracking robustness by combining RGB and thermal infrared (TIR) modalities, addressing limitations of RGB-only trackers under challenging conditions such as low light and appearance variations. However, most existing RGB-T trackers rely on complex fusion modules or modality-specific architectures, sacrificing efficiency for performance. In this paper, we propose a novel Multi-level Self-Distillation (MSD) framework that adapts a one-stream RGB tracker to the RGB-T setting without modifying the network architecture or adding any extra parameters. RGB and TIR inputs are jointly processed through a shared backbone, and training is guided by a combination of self-supervised and supervised objectives to enhance cross-modal feature representation. The self-supervised component includes a contrastive loss that aligns semantically consistent regions across template-search pairs, as well as a modality-gap alignment loss that reduces discrepancies between RGB and TIR features. These internal signals complement task-driven supervision, including an intermediate focal loss that strengthens early localization by enhancing shallow and mid-level features, modality-specific losses that preserve distinctive cues under partial modality degradation, and a fused tracking loss that drives final bounding box prediction. Comprehensive evaluations on LasHeR, RGBT234, and GTOT benchmarks demonstrate that MSD achieves state-of-the-art tracking accuracy while maintaining the computational efficiency of the original RGB tracker. Our work establishes a new paradigm in multi-modal tracking by demonstrating that optimized training strategies can outperform complex architectural modifications, offering significant practical advantages for real-world deployment.

Deep neural networks using generative diffusion prior have provided the state-of-the-art performances for the task of blind image super resolution. Thanks to their powerful image generation capability, these deep networks are able to produce high-quality visual signals with realistic textures and structures. However, since these schemes employ a very large number of parameters, their training process is often difficult, and therefore, their performances can be limited. In order to address this, in this paper, we propose a diffusion-based blind image super resolution scheme, which by using a novel learning algorithm with invertible neural networks, is able to provide superior results. Specifically, we argue that because of the reversibility property of invertible neural networks, they are able to generate degraded low-quality images, whose super resolved versions are the upper bound of the image super resolution function space. The inclusion of such visual signals in the training process of our blind image super resolution network leads to facilitating the learning paradigm and achieving higher performances. We show that our proposed blind image super resolution scheme is able to outperform the state-of-the-art methods.

Convolutional neural network (CNN) pruning has traditionally relied on heuristically designed importance criteria, often leading to limited generalizability and inconsistent performance. In this article, we propose a novel framework centered around filter replacement (FR), introducing pruning as a process of replacing selected filters with zero filters. Through a rigorous analysis, we derive an upper bound on the absolute error in the output of the subsequent layer and use this bound to define an efficient importance function. This importance function exhibits $\gamma $ -weakly submodular properties, enabling the development of a simple, low-complexity, and data-free oblivious algorithm for selecting filters to prune. In addition, we extend the FR framework to include nonzero filter alternatives, leveraging a best-approximation technique to construct optimal replacements for the pruned filters. Extensive experiments on benchmark networks and datasets validate the effectiveness of our method. The proposed approach achieves state-of-the-art results, with a complexity comparable to basic techniques such as $l_{2}$ -norm pruning. Notably, our pruning method achieves 76.52% accuracy (ACC) in ResNet-50 on the ImageNet dataset, surpassing the baseline of 75.15%, while reducing network parameters by 25.5%. Our proposed resource efficiency (RE) metric assesses that the layer interdependence-aware pruning (LIAP) method is up to $10^{11}$ times more efficient than existing techniques, setting a new standard for resource-aware CNN pruning.

RGB-T object tracking leverages visible and thermal infrared modalities to enhance robustness in challenging environments. While deep learning-based RGB-T trackers predominantly use feature-level fusion, pixel-level fusion remains underexplored. This paper introduces the Hierarchical Feature Difference Auto-Encoder (HFDAE), a novel pixel-level fusion approach that refines the RGB modality before tracker input. HFDAE adaptively enhances RGB content using hierarchical TIR features, dynamically emphasizing object saliency. HFDAE consists of three key components: (1) a shallow RGB autoencoder that preserves structural and color information, (2) a TIR encoder with variable-depth decoders generating hierarchical TIR representations, and (3) a fusion module that integrates salient thermal features into the RGB image. Salient features are extracted by computing differences between hierarchical and base TIR images, which are then added to the base RGB image to generate the final fused output. Unlike conventional pixel-level fusion methods, HFDAE is optimized directly for tracking, learning fusion strategies without predefined modality assumptions. Extensive experiments on benchmark datasets demonstrate HFDAE’s superior tracking accuracy and robustness across diverse scenarios. The proposed approach improves the base tracker’s precision rate by approximately 10%. Code is available at https://github.com/mohamed-e-awad/HFDAE.

The field of single-image super-resolution (SISR) has seen significant advancements with the emergence of deep convolutional neural networks, where residual learning techniques have contributed to notable improvements in reconstruction quality. Among these approaches, SwinIR [1], a Transformer-based model, has demonstrated impressive performance by leveraging hierarchical self-attention mechanisms to capture both local fine-grained structures and global contextual dependencies. However, improving image quality while maintaining computational efficiency remains a key challenge. To address this, we propose a multi-scale SwinIR inception-based network, an enhanced SISR framework that draws inspiration from the Inception module to refine feature extraction across multiple scales without introducing significant computational overhead due to the complexity of the network architecture. Instead of directly implementing the Inception module, we adopt its core idea of parallel multi-scale processing, where multiple convolutional layers with different receptive fields operate simultaneously to extract spatial features at varying scales. This strategic enhancement significantly improves PSNR over the original SwinIR model while increasing the parameter count by only 65K. Our model integrates hierarchical self-attention with multiscale feature extraction to strengthen the representation of structural details in low-resolution images. Experimental results demonstrate that EMS(Enhanced multi-scale network) consistently outperforms state-of-the-art SISR models across multiple benchmark datasets, delivering improved visual fidelity and superior quantitative performance without a significant increase in computational complexity.

This work proposes a novel multi-scale attention-enhanced dual-task network, MSA-DTNet, to simultaneously address two critical tasks, segmentation and registration. MSA-DTNet is designed for high-resolution 3D MRI data to incorporate multi-scale convolutions to capture both local and global features and attention mechanisms to enhance the model’s focus on key anatomical structures. By jointly optimizing segmentation and registration tasks, our network improves anatomical consistency and overall performance in medical image processing. The segmentation decoder produces high-quality segmentation maps, while the registration decoder outputs a displacement field for aligning images with a reference. A novel hybrid loss function is also proposed to optimize the model during training. The experiments on the brain MRI dataset demonstrate that MSA-DTNet outperforms existing state-of-the-art networks in terms of dice score (DSC), intersection over union (IoU), precision and recall in segmentation, and DSC and mean squared error (MSE) in registration tasks. Our model also achieves significant performance improvements, even with limited labeled data, by leveraging semi-supervised learning.

In this paper, we introduce Adaptive Multi-Scale Spatiogram (AMS), a new feature descriptor that aims to enhance both robustness and accuracy in feature matching, under various transformations such as viewpoint, scale, rotation, and illumination changes. The proposed method dynamically adapts to local image features by capturing complex details in texture-rich areas while simplifying representation in smoother regions. The method combines a Bayesian model for detecting keypoints, adaptive binning based on entropy, and a low-rank approximation to handle a wide variety of scene changes. Experimental results demonstrate that the proposed method outperforms traditional descriptors in complex transformations, though with a moderate computational cost. The overall results suggest that AMS achieves a satisfactory balance between effectiveness and efficiency, making it suitable for advanced computer vision applications.

Diffusion models have provided the state-of-the-art performances for different computer vision tasks, including the task of underwater image enhancement. One of the challenges in the task of underwater image enhancement is that various spatial regions of the image require different restoration techniques. In order to address this, we propose a novel diffusion-based underwater image enhancement network, in which by employing the two ideas of uncertainty-aware learning and feature recalibration based on the color tones dominated in the underwater environments, it is able to provide superior performances. Specifically, the former idea strives to process various spatial regions of the underwater image based on their restoration uncertainty, while the latter technique recalibrates the features generated by the diffusion model by taking the various color tones in the underwater environments into consideration. The results of different experimentations show the superiority of the proposed diffusion-based model over the other state-of-the-art underwater image enhancement networks.

Recent advancements in the classification of Alzheimer's disease have leveraged the automatic feature generation capability of convolutional neural networks (CNNs) using neuroimaging biomarkers. However, most of the existing CNN-based methods often disregard the local features of the brain data, which leads to a loss of subtle fine-grained features in the brain imaging data. Moreover, the existing CNN architectures, which mainly rely on global features, do not pay much attention to the discriminability of the extracted features for the task of classification of Alzheimer's disease. Moreover, the existing architectures often end up using a large number of parameters to enhance the richness of the extracted features. This paper proposes a novel lightweight deep CNN, which extracts local and global contextual features from the sagittal slices of structural MRI data and uses both of these two types of features for the classification of the disease. The main idea used in designing the proposed network is to process separately the local and global features by using modules that pay a special attention to extract local and global contextual features. The fused local and global contextual features are then used for the classification of Alzheimer's disease. The proposed network is tested for the binary and multiclass classifications of the disease using the MR images taken from the ADNI database. The proposed network is shown to provide a performance that is significantly higher than that provided by other existing state-of-the-art networks, yet using a number of parameters that is a small fraction of that used by the other schemes.

The brain-computer interface (BCI) establishes a connection between a device and the human brain, with electroencephalography (EEG) signal is being used as the most common means for such a communication. We use EEG signal data that has a very limited number of samples for the motor imagery (MI) classification task. This paper proposes a novel densely connected residual graph convolutional network (DenseResGCN) and uses it in developing a few-shot learning method called FewShotEEG method. Our proposed method is capable of classifying the limited EEG signal data into four MI classes. The proposed method outperforms the state-of-the-arts few-shot methods in terms of the accuracy.

In this paper, we propose a novel convolutional neural network called MAGNet that employs multi-scale and global attention mechanisms for the task of medical image segmentation. This network is shown effectively to handle the segmentation task of an image of a given modality provided the network is suitably trained using a training set of the same modality. Experiments are performed to train the proposed network using three different training sets of images (CT, colonoscopy, and non-mydriatic 3CCD images), each acquired from a different imaging technique, resulting in three different trained models. The three trained models are tested on the respective test sets. Each model is shown to significantly outperform the state-of-the-art networks in terms of intersection over union, dice coefficient, and accuracy.

Obtaining robust representational features is essential for achieving high performance in image retrieval. One effective strategy to boost the representational capacity of deep features is through feature aggregation, a technique that combines features from various feature maps of a deep network. In this paper, we propose a novel feature aggregation method that integrates feature maps from different levels of abstraction within a deep network into a single feature vector. The proposed method employs the particle swarm optimization algorithm to adaptively assign optimal weights to each group of feature maps. Extensive experiments are conducted to validate the effectiveness of our proposed feature aggregation method. It is shown that the proposed method significantly outperforms existing state-of-the-art feature aggregation methods on various benchmark datasets.

JPEG is one of the most popular image compression techniques , with numerous applications ranging from medical imaging to surveillance systems. Since JPEG introduces the blocking artifacts to the decompressed visual signals, enhancing the quality of these images is of paramount importance . Recently, various deep neural networks have been proposed for JPEG image deblocking that can effectively reduce the blocking artifacts produced by the JPEG compression technique. However, most of these schemes could only handle decompressed images generated by a set of specific JPEG quality factor (QF) values employed in the network training process. Therefore, when the images are obtained by the JPEG QF values other than those used in the network training process, the performance of deep learning-based JPEG image deblocking schemes drops significantly. To address this, in this paper, we propose a novel deep learning-based blind JPEG image deblocking method, which employs out-of-distribution detection to perform deblocking efficiently for various quality factor (QF) values. The proposed scheme can distinguish between the decompressed images using the QF values used in the training set and those using the QF values not used in the training set, and then, a suitable deblocking strategy for generating high-quality images is developed. The proposed scheme is shown to outperform the state-of-the-art JPEG image deblocking methods for various QF values.

Image retrieval aims to find similar images to a given query by matching features extracted directly from the images of a database. Deep convolutional neural networks provide an excellent framework for obtaining highly representative feature vectors from images to improve an image retrieval method. Deep residual networks perform better than existing deep networks, as they can incorporate useful information into the feature vectors through residual learning by designing appropriate operations in the employed residual block. One type of such information is spatial information obtained at different scales and levels of abstraction. In this paper, a novel residual block is proposed to generate a rich set of features for the task of image retrieval. The development of the residual block consists of three modules: a hierarchical spatial feature extraction module focusing on spatial information at different abstraction levels, a multi-scale feature extraction module that generates features at three different scales, and a feature fusion module. The results of experiments on various datasets and an ablation study show that the proposed residual block noticeably improves the representational capacity of the network, which, in turn, significantly enhances the retrieval performance of the deep image retrieval network.

The representational quality of the generated feature vectors for images is essential for image retrieval models to achieve high performance. Spatial information is crucial in obtaining highly representative feature vectors for image retrieval, and deep convolutional neural networks provide an excellent framework to generate such features. Through convolutional operations, deep convolutional neural networks include spatial information in the feature maps. However, most available architectures cannot include adequate spatial details in the feature maps required for high-performance image retrieval. Deep residual networks are deep networks capable of including useful information through residual learning. This paper proposes a novel residual block to generate feature maps by focusing on spatial information. The proposed residual block comprises three modules: a spatial feature extraction module, a hierarchical feature extraction module, and a feature fusion module. The first module includes spatial information in the feature maps at different levels of abstraction, while the second module includes spatial information using conventional convolution hierarchy. The third model fuses the outputs of the first two modules to provide a very rich set of feature maps. The present study tests a deep network employing the proposed residual block. The results indicate that the proposed network performs comparably or is superior to state-of-the-art methods on standard benchmarks, thus showing the effectiveness of the proposed residual block in improving the representational capacity.

High-frequency components are the most crucial parts of the visual signals for the task of image super resolution. The deep image super resolution networks that are able to process the high-frequency components efficiently can provide high performances. In view of this, in this paper, we develop a new residual block for image super resolution, in which the feature attention process is carried out by focusing on various high-frequency components of the feature tensors. Specifically, we design a novel multi-dimensional filter design technique for the task of image super resolution, and employ it for obtaining a finite impulse response (FIR) high-pass filter bank to be embedded in a deep super resolution network for the feature attention process. Moreover, we utilize two other feature attention processes in the proposed residual block, namely, multi-scale transformerbased and convolutional learnable feature attention mechanisms, to generate rich sets of feature maps for a deep super resolution network. The results of different experiments demonstrate the effectiveness of the various modules of the proposed residual block in enhancing the super resolution performance

Feature attention is a technique used in deep neural networks to provide a discriminative processing of the various regions in an image based on their significance for enhancing the image restoration performance. In this paper, we develop a novel image restoration network, in which the feature maps extracted by the network are recalibrated using a pixel-wise feature attention and the recalibration process is guided by the structural and textural information of the image resulting from the Otsu’s method for its segmentation. It is shown that using this segmentation guidance strategy for recalibrating feature maps is indeed helpful in enhancing the quality of the restored images. The proposed image restoration network outperforms the state-of-the-art light-weight image restoration networks on benchmark datasets.

Alzheimer’s disease (AD) is a progressive brain disorder affecting millions of people worldwide. An accurate diagnosis of AD plays a significant role in identifying the progression of the disease at its prodromal stage, i.e., mild cognitive impairment (MCI). In this paper, we propose a lightweight deep model to classify the patients into diagnostic groups, AD vs. normal control (NC) or progressive MCI (pMCI)vs. stable MCI (sMCI), with high accuracy, using MRI data. The proposed model uses separable and attention-based convolution operations. The separable convolution can reduce the complexity of the model by splitting a kernel into two separate kernels that do depth-wise and pointwise convolution operations, respectively. Moreover, integrating an attention-based convolution, which concatenates the convolutional and attentional feature maps, can capture the most relevant features for improved classification with fewer filters. From the experimental results on the Alzheimer’s disease neuroimaging initiative (ADNI) database, compared to the state-of-the-art methods, it is observed that the proposed method shows significant improvement in the classification performance in terms of accuracy, specificity, sensitivity, and AUC. In addition, the proposed method drastically reduces the number of parameters without affecting the performance.

Deep light-weight image super resolution networks that provide a high performance have numerous real-life applications, such as mobile devices and multimedia systems. Hence, analyzing the capability of such deep networks in providing a similar performance between the cases that they are applied to the images with and without distributions similar to that of the training is crucial. In this paper, we carry out the robustness analysis of the deep state-of-the-art light-weight super resolution networks by proposing and using three metrics that are based on the statistical information of the super resolved images in both pixel level and feature level. The results of our metrics for the deep state-of-the-art light-weight super resolution networks demonstrate the behavior of such networks against realistic distribution shift in the test dataset.

ThiNet is a recent method for pruning convolutional neural networks. This method uses a norm of a subset of the components of the output resulting from the convolutional layer succeeding the layer from which the filters are to be removed for pruning the network. The ThiNet algorithm is very time-consuming, in view of the fact that the filters for removal are selected one by one iteratively. In this paper, we propose a modified version of ThiNet, in which the same information on the output of the same convolutional layer as used by ThiNet is employed to select all the filters together in a single step, for pruning the network. The proposed modified algorithm is shown to have a time-complexity that is only a small fraction of that of ThiNet or any other state-of-the-art algorithm and that the pruned network has almost the same reduction in its accuracy as that of the network pruned by ThiNet.

Morphological operations are nonlinear mathematical operations that are capable of performing signal processing tasks based on the structures and textures of the signals. With this motivation of the capability of morphological operations, in this paper, a novel residual block that can generate morphological features of images and fuse them with the conventional hierarchical features has been proposed. The proposed residual block is then used to design a light-weight deep neural network architecture in a residual framework for the task of image super resolution. It is shown that a fusion of morphological features of images with the conventional hierarchical features can improve the super resolution capability of a deep convolutional network. Experiments are performed to demonstrate the effectiveness of the proposed idea of using morphological operations and the superiority of the network designed based on this idea in super resolving low quality images.

The design of ultralight-weight super-resolution convolutional neural networks capable of providing images with high visual quality is crucial in many real-world applications with limited power and storage capacity, such as mobile devices and portable cameras. In this paper, a new ultralight-weight super-resolution network, based on the idea of using multiresolution level feature interpolation in a residual framework, is developed. In the proposed network, the multiple resolution level interpolated features generated are fused and the resulting feature maps are added to the residual features obtained from a shallow convolutional neural network. The proposed network is applied to various benchmark datasets and is shown to outperform the state-of-the-art ultralight-weight image super-resolution networks existing in the literature.

Brain computer interface is a developing technology that can provide enhanced quality of life to individuals suffering from various disabilities. In this work, a new binary electroencephalography (EEG) signal decoding algorithm is proposed using a modified common spatial pattern and capsule network. The proposed method is realized by extracting the spectral-temporal common spatial pattern features from the EEG signals while preserving the time resolution of the signal. The resulting features are fed into the capsule network for automatic feature extraction and classification. The capsule network is known to be superior to convolutional neural networks in requiring less training data, which makes it a promising candidate for EEG signals classification. The performance of the proposed method is evaluated and compared to that of the other methods by conducting several experiments. The results demonstrate that the proposed method provides recognition accuracy higher than that provided by other methods.

This paper presents a novel two-step method for channel estimation in millimeter wave hybrid MIMO systems over doubly (time and frequency) selective channels. In the first step, the dominant delay taps of the channel are selected by using an energy detector. In the second step, utilizing the basis expansion model (BEM) for effectively representing doubly selective channels, a BEM-based block orthogonal matching pursuit algorithm is proposed for estimating the gains of the selected delay taps. The proposed method benefits from exploiting the entire available training sequence to estimate all channel parameters, is capable of capturing channel variations across the whole training frames and does not require any feedback. Computer simulations and complexity evaluations show that the proposed approach can significantly improve the mean squared error performance compared with the existing techniques without increasing the computational complexity.

In this paper, a new regularization term in the form of L1-norm based fractional gradient vector flow (LF-GGVF) is presented for the task of image denoising. A fractional order variational method is formulated, which is then utilized for estimating the proposed LF-GGVF. Overlapping group sparsity along with LF-GGVF is used as priors in image denoising optimization framework. The Riemann-Liouville derivative is used for approximating the fractional order derivatives present in the optimization framework. Its role in the framework helps in boosting the denoising performance. The numerical optimization is performed in an alternating manner using the well-known alternating direction method of multipliers (ADMM) and split Bregman techniques. The resulting system of linear equations is then solved using an efficient numerical scheme. A variety of simulated data that includes test images contaminated by additive white Gaussian noise are used for experimental validation. The results of numerical solutions obtained from experimental work demonstrate that the performance of the proposed approach in terms of noise suppression and edge preservation is better when compared with that of several other methods.

To better understand early brain development in health and disorder, it is critical to accurately segment infant brain magnetic resonance (MR) images into white matter (WM), gray matter (GM), and cerebrospinal fluid (CSF). Deep learning-based methods have achieved state-of-the-art performance; h owever, one of the major limitations is that the learning-based methods may suffer from the multi-site issue, that is, the models trained on a dataset from one site may not be applicable to the datasets acquired from other sites with different imaging protocols/scanners. To promote methodological development in the community, the iSeg-2019 challenge (http://iseg2019.web.unc.edu) provides a set of 6-month infant subjects from multiple sites with different protocols/scanners for the participating methods. T raining/validation subjects are from UNC (MAP) and testing subjects are from UNC/UMN (BCP), Stanford University, and Emory University. By the time of writing, there are 30 automatic segmentation methods participated in the iSeg-2019. In this article, 8 top-ranked methods were reviewed by detailing their pipelines/implementations, presenting experimental results, and evaluating performance across different sites in terms of whole brain, regions of interest, and gyral landmark curves. We further pointed out their limitations and possible directions for addressing the multi-site issue. We find that multi-site consistency is still an open issue. We hope that the multi-site dataset in the iSeg-2019 and this review article will attract more researchers to address the challenging and critical multi-site issue in practice.

The JPEG compression scheme introduces blocking artifacts when the images are decompressed. JPEG image deblocking schemes based on deep neural networks map a JPEG decompressed image to its corresponding deblocked image. Employing a residual block that is capable of generating a rich set of high frequency residual features in a deep JPEG image deblocking network can improve its representational capability, and therefore, enhance the network performance. In this paper, we propose two residual blocks that generate rich high frequency residual features. The first residual block generates features from the high frequency component of its input signal in addition to generating conventional hierarchical residual features using convolutional operations. The second one is a fractal residual block that is developed by replacing the conventional convolutions in the first block by the block itself. The two proposed residual blocks are, respectively, used in recursive (non-fractal) and non-recursive (fractal) neural networks for the task of JPEG deblocking. The results of the experiments performed on the two proposed deblocking networks show their performance superiority over the respective state-of-the-art deblocking networks.

Residual blocks use skip connections in order to facilitate the flow of information in the network and thus, provide a good network performance. As different objects in a generic image appear at different scales, employing a multi-scale feature generation module in a residual block for image super resolution can further improve the network performance. In this paper, a new residual block that generates features at multiple scales is proposed for the task of image super resolution. In order to enhance the representational capability of the network while keeping its complexity low, the proposed residual block uses two different feature generation techniques, namely, multi-scale granular channel feature generation and uni-scale holistic channel feature generation, and fuses their output feature maps. It is shown that the network using the proposed residual block outperforms the state-of-the-art lightweight super resolution networks on four benchmark datasets with various scaling factors.

Deep neural networks use a nonlinear end-to-end mapping in order to transform a low resolution image to the high resolution one. Residual blocks facilitate the flow of the information in deep neural networks and enhance the network performance. In this paper, a new residual block that enhances the representational capability of a super resolution network is proposed. The proposed residual block combines the features generated in various receptive fields using different hierarchical levels of convolution operations or convolution operations in conjunction with the space-to-depth and depth-to-space operations in order to provide a rich set of residual features. The experimental results demonstrate the superiority of the super resolution network using the proposed residual block over the state-of-the-art light-weight super resolution networks in terms of objective and subjective metrics.

Deep convolutional networks provide very high quality super resolution images through a learning process by a nonlinear end-to-end mapping between low and high resolution images. Many of the state-of-the-art super resolution networks employ residual blocks in their network architectures, where in each residual block the high frequency residual signals are added to the feature maps input to the block. In this paper, a new residual block is proposed for the problem of image super resolution. The proposed residual block consists of three modules, namely, feature transformation module, nonlinear edge extraction module and feature fusion module. The feature transformation module produces high frequency residual signals and the nonlinear edge extraction module extracts the edges of the features input to the block. These generated high frequency features are then fused using the feature fusion module in order to produce a very rich set of high frequency residual features. The performance of the super resolution network using the proposed residual block is compared with that of the state-of-the-art light-weight super resolution schemes on four benchmark datasets. It is shown that the proposed super resolution scheme outperforms the state-of-the-art light-weight super resolution networks, when both the performance and number of parameters of the network are simultaneously taken into consideration.

Deep image super resolution networks use a nonlinear end-to-end mapping between the low and high resolution versions of an image and therefore, provide a good performance. As the different parts of a single image appear in different scales, developing a deep learning based image super resolution scheme that is capable of generating features at different scales and levels is essential. In this paper, a new residual block is proposed with a view of generating a rich set of features extracted at different scales and levels. The development of the proposed block is carried out using two distinct strategies, the first one focussing on generating features directly in two different scales, whereas the second one aims at generating multi-scale features indirectly by extracting them from two different hierarchical levels of abstraction. It is shown through experimental results that the proposed scheme of designing the residual block results in a network that provides a superior performance with reduced number of parameters than that provided by the light-weight networks using other types of residual blocks.

Restoration of a blurred and subsampled image is an ill-posed problem. In this paper, a two-stage convolutional network is proposed to carry out the processes of upsampling and deblurring to restore the original image. The main idea in the proposed scheme is that the deblurring process is attempted on a high PSNR image obtained after removing the ringing effect that is necessarily caused by the upsampling process. The evaluation of the proposed scheme is carried out using a benchmark dataset in terms of PSNR. The scheme is shown to outperform the state-of-the-art schemes, namely, the sparse coding network, the non-local means filters and the centralized sparse representation.

Image compression is a field in image processing that tries to remove redundant information in an image. Losing the information in lossy compression techniques such as JPEG results to artifacts in the decompressed image that necessitates the image restoration. In this work, a new image restoration scheme based on deep neural nets and maxout activation functions for the application of image deblocking is proposed. Experimental results are presented to demonstrate the superiority of the proposed method both in terms of subjective and objective metrics.

Salient object detection is an active research topic due to several potential applications in image compression, scene understanding, image retrieval, and so forth. In this paper, a salient object detection method is proposed by leveraging the recent advances in graph signal processing. Since, the image boundary regions generally belong to the image background, a distribution-based boundary contrast map is generated. Also, the graph representation of the image is used to compute the connectivity of the image regions to the image boundary as well as those to their local neighbors and the image foreground. The connectivity maps obtained are fused with the boundary contrast map in order to obtain the image saliency map. Several experiments are conducted to evaluate the performance of the proposed salient object detection method and to compare it with the state-of-the-arts. Results on datasets of images demonstrate that the proposed method achieves superior performance to the state-of-the-art methods in terms of precision, recall, and mean absolute error values.

In this paper, a new scheme for single image super resolution using convolutional neural networks and subband reconstruction theory is proposed. In the design of the network, which is referred to as SRSubBandNet, each subband of the residual signal between the high and low resolution images is reconstructed from all the previous subbands. Skip connections between the first, middle and the last SRBs are utilized to address the gradient vanishing problem in the proposed network. SRSubBandNet provides competitive results in terms of both subjective and objective qualities when applied to various benchmark datasets.

Busy life as well as the prevalence of infotainment is increasingly making people more occupied even during tasks that require serious attention. One such task is driving and at the same time getting involved in activities that may distract drivers cognitively from watching the road and cause fatal accidents. This paper presents a method that is capable of monitoring different types of distractions, such as talking and texting on cell phone, casual eating, and operating cabin equipment while driving, so that a driver can be assisted to remain cautious on the road. The proposed method automatically detects and tracks fiducial body parts of a driver from video captured by a camera mounted on the front windshield inside a vehicle. Relative distances between the tracking trajectories are used as features that represent actions of the driver. Then, the well-known kernel support vector machine is applied for recognizing a particular distraction from the features extracted from body parts. The proposed feature is also compared with previously employed features for tracking-based human action recognition schemes to substantiate its better result in terms of mean accuracy and robustness for distraction recognition. The effectiveness of the proposed method of distraction recognition is also analyzed with respect to tracking errors.

Structural information, in particular, the edges present in an image are the most important part that get noticed by human eyes. Therefore, it is important to denoise this information effectively for better visualization. Recently, research work has been carried out to characterize the structural information into plain and edge patches and denoise them separately. However, the information about the geometrical orientation of the edges are not considered leading to sub-optimal denoising results. This has motivated us to introduce in this paper an adaptive steerable total variation regularizer (ASTV) based on geometric moments. The proposed ASTV regularizer is capable of denoising the edges based on their geometrical orientation, thus boosting the denoising performance. Further, earlier works exploited the sparsity of the natural images in DCT and wavelet domains which help in improving the denoising performance. Based on this observation, we introduce the sparsity of an image in orthogonal moment domain, in particular, the Tchebichef moment. Then, we propose a new sparse regularizer, which is a combination of the Tchebichef moment and ASTVbased regularizers. The overall denoising framework is optimized using split Bregman-based multivariable minimization technique. Experimental results demonstrate the competitiveness of the proposed method with the existing ones in terms of both the objective and subjective image qualities.

Sleep apnea, a serious sleep disorder affecting a large population, causes disruptions in breathing during sleep. In this paper, an automatic apnea detection scheme is proposed using single lead electroencephalography (EEG) signal to discriminate apnea patients and healthy subjects as well as to deal with the difficult task of classifying apnea and nonapnea events of an apnea patient. A unique multiband subframe based feature extraction scheme is developed to capture the feature variation pattern within a frame of EEG data, which is shown to exhibit significantly different characteristics in apnea and nonapnea frames. Such within-frame feature variation can be better represented by some statistical measures and characteristic probability density functions. It is found that use of Rician model parameters along with some statistical measures can offer very robust feature qualities in terms of standard performance criteria, such as Bhattacharyya distance and geometric separability index. For the purpose of classification, proposed features are used in K Nearest Neighbor classifier. From extensive experimentations and analysis on three different publicly available databases it is found that the proposed method offers superior classification performance in terms of sensitivity, specificity, and accuracy.

Data security is a main concern in everyday data transmissions over the Internet. A possible solution to guarantee secure and legitimate transaction is via hiding a piece of tractable information into the multimedia signal, that is, watermarking. In this paper, we propose a new color image watermarking scheme and its corresponding detector in the sparse domain. The watermark detector aims at verifying the ownership and circumventing any unauthorized duplication of the digital data. Most of the existing color image watermarking schemes disregard the inter-channel dependencies. In view of this, we take into account the interchannel dependencies between RGB channels and interscale dependencies of the sparse coefficients of color images by employing the hidden Markov model. An efficient detector is designed by establishing a binary hypothesis test through which the existence of the hidden watermark is examined. Experiments are conducted to evaluate the performance of the proposed watermark detector for color images. The results show that the proposed detector provides detection rates higher than those provided by the other detectors, even in the presence of attacks. It is also shown that the proposed detector exhibits better performance in terms of the robustness of the embedded watermark.

In this paper, residual redundancy in compressed videos is exploited to alleviate transmission errors using joint source channel arithmetic decoding. A new method is proposed to estimate a priori probability in MAP metric of H.264 intra modes decoder. The decoder generates a decoding tree using a breadth first search algorithm. An introduced statistical model is then implemented stage by stage over the decoding tree. In this model, a priori PMF of intra block modes in a macroblock is estimated from the intra block modes seated in its spatially adjacent macroblocks previously generated up to the current stage of the decoding tree. The estimated PMFs are categorized as either reliable or unreliable based on their local entropies. In the unreliable case, the decoder assumes uniform PMF and switch to ML metric instead. The simulation results show the proposed method reduces the error rate 1 % to 13% at various SNRs compared to the ML.

In this paper, first, a new fusion technique, referred to as hybrid fusion (HBF) technique, based on feature-level fusion and the best unimodal system for multimodal biometric system recognition, is proposed. Secondly, a new weighting technique, referred to as mean-extrema based confidence weighting (MEBCW) technique, based on the scores obtained from feature-level fusion and the best unimodal system, is proposed. Finally, a weighted hybrid fusion, referred to as weighted hybrid fusion (WHBF) technique, is developed by incorporating MEBCW in HBF, in order to improve the overall recognition rate of a multimodal biometric system. The performance of the proposed method, in terms of equal error rate and genuine acceptance rates @5.3% and @7.2% false acceptance rates, is evaluated on a multi-biometric system. The experimental results show that the performance of a multi-biometric systems using the proposed fusions is superior to that of the uni-biometric systems or to that of the system using existing level of fusions.

The human visual system is attracted to the most dominant part of the image which is called salient region. There has been a surge of interest in the past few years to efficiently detect the salient regions of images. In this study, a new salient region detection method is proposed using the non‐subsampled contourlet transform. It is known that this transform is capable of providing a multiscale, multi‐directional and translation invariant decomposition of images. The proposed saliency detection method is realised by extracting various local and global features from the non‐subsampled contourlet coefficients of the colour channels. A saliency map is obtained based on a linear combination of the local features and the distribution of the global features. In order to provide a better preservation of the structure and boundary of the objects and to obtain a more uniformly highlighted salient region, the saliency map is abstracted using an optimisation framework. Several experiments are conducted on sets of natural images to evaluate the performance of the proposed method. The results show that the performance of the proposed method is superior to that of the other existing methods in terms of precision‐recall performance, F ‐measure, and mean absolute error values.

The vector-based hidden Markov model (HMM) is a powerful statistical model for characterizing the distribution of the wavelet coefficients, since it is capable of capturing the subband marginal distribution as well as the inter-scale and cross-orientation dependencies of the wavelet coefficients. In this paper we propose a scheme for designing a blind multibit watermark decoder incorporating the vector-based HMM in wavelet domain. The decoder is designed based on the maximum likelihood criterion. A closed-form expression is derived for the bit error rate and validated experimentally with Monte Carlo simulations. The performance of the proposed watermark detector is evaluated using a set of standard test images and shown to outperform the decoders designed based on the Cauchy or generalized Gaussian distributions without or with attacks. It is also shown that the proposed decoder is more robust against various kinds of attacks compared with the state-of-the-art methods.

The performance of a multi-biometric system can be improved using an efficient normalization technique under the simple sum-rule-based score-level fusion. It can also be further improved using normalization techniques along with a weighting method under the weighted sum-rule-based score-level fusion. In this paper, at first, we present two anchored score normalization techniques based on the genuine and impostor scores. Specifically, the proposed normalization techniques utilize the information of the overlap region between the genuine and impostor scores and their neighbors. Second, we propose a weighting technique that is based on the confidence of the matching scores by considering the mean-to-maximum of genuine scores and mean-to-minimum of impostor scores. A multi-biometric system having three biometric traits, fingerprint, palmprint, and earprint, is utilized to evaluate the performance of the proposed techniques. The performance of the multi-biometric system is evaluated in terms of the equal error rate and genuine acceptance rate @0.5% false acceptance rate. The receiver operating characteristics are also plotted in terms of the genuine acceptance rate as a function of the false acceptance rate.

A salient region is part of the image that captures the greatest attention by the human visual system. In this paper, we propose a novel salient region detection technique in the non-subsampled contourlet domain. The image is first decomposed into non-overlapping patches in order to fully exploit the repetitive patterns in the image. It is known that the non-subsampled contourlet transform provides an efficient multi-resolution, multi-directional, localized and shift invariant decomposition of images. In view of this, by using the statistical properties of non-subsampled contourlet coefficients of image patches, a set of feature descriptors are extracted to construct the feature map for each color channel. An entropy-based criterion is proposed to combine the channel feature maps into a saliency map. Simulations are conducted on a dataset of natural images to evaluate the performance of the proposed method and to compare it with that of the other existing methods. The results show that the proposed salient region detection method provides higher precision, recall, and F-measure and lower mean absolute error values as compared to the other existing methods.

Multimedia data piracy in the Internet is a growing problem, since it provides easy and fast data transmission. Watermarking is regarded as a solution to restrain unauthorized duplication or distribution data. Image watermarking research mostly focuses on grayscale images with an extension to color images. However, most of these techniques ignore dependencies between color channels. In view of this, in this work, a multichannel color image watermarking technique and its corresponding detector in the wavelet domain is proposed. The inter-channel dependencies between RGB channels and inter-scale dependencies of the wavelet coefficients of color image are taken into account by employing the vector-based hidden Markov model. We conduct experiment on a set of color images to assess the performance of the proposed watermark detector. The results show that the performance of the proposed detector is superior to that of the other detectors in terms of the imperceptibility of the embedded watermark and the detection rate. It is also shown that the proposed detector has better performance in presence or absence of different kinds of attacks in comparison to the other existing methods.

Statistical modeling of high dimensional, correlated, and complex imaging data obtained from various longitudinal neuroimaging studies has become an inevitable part of automatic disease diagnosing tasks. In this paper, we propose a parametric brain image modeling based on the statistical properties of non-subsampled shearlet transform (NSST) coefficients. The NSST detail coefficients of multimodal neuroimaging data exhibit highly non-Gaussian property, i.e., the probability density function (PDF) of the NSST coefficients are sharply peaked around zero with heavy tails. As a consequence, the marginal statistics of the detail subband coefficients are modeled by student's t location-scale PDF, which has heavier tails (more prone to outliers) than the Gaussian distribution for smaller values of the shape parameter. The Jensen-Shannon divergence (JSD) goodness-of-fit shows that the detail NSST subbands of neuroimaging data in the longitudinal Alzheimer's disease neuroimaging initiative database are well approximated by student's t location-scale distribution compared to that by the traditional generalized Gaussian distribution.

This paper investigates the spectral efficiency of multiuser multiple-input multiple-output systems with a large number of antennas at the base station that serves single-antenna users in one cell. It is assumed that the base station estimates the channel with the help of uplink training and then employs the zero-forcing technique to detect the data signals transmitted by the various users. An optimal power control scheme over pilot and data power based on large-scale fading is proposed to maximize the sum spectral efficiency for a given total energy budget in a coherence interval. Simulation results show that the spectral efficiency of the proposed method is superior to that of other existing methods. It Is also shown that, In order to maximize the sum spectral efficiency, more power should be allocated to the data signal power at high signal-to-nolse ratios and less power at low signal-to-noise ratios.

This paper studies the problem of maximizing the spectral efficiency in a massive multi-user MIMO downlink (DL) system where the base station employs a large number of antennas and serves single-antenna users in one cell, assuming time- division duplex transmission. In order to maximize the spectral efficiency in a DL transmission, a new method for power allocation among users is proposed based on the water-filling approach. To this end, a lower bound on the achievable rate is derived for each user in the presence of maximum- ratio transmission precoding and employed in defining the spectral efficiency. Since maximizing the spectral efficiency is an NP-hard problem, an effective algorithm is also proposed to find Karush-Kuhn-Tucker (local maximum) points. The performance of the proposed power allocation method is verified by conducting simulations and shown to be superior to other existing methods in terms of higher spectral efficiency.

In this paper, a noise robust formant frequency estimation scheme is developed based on a spectral model matching algorithm. Considering the vocal tract as an autoregressive system, a spectral model of repeated autocorrelation function (RACF) of band-limited speech signal is proposed. It is shown that because of the repeated autocorrelation operation on band-limited signal, the proposed model can exhibit prominent formant characteristics. First from given noisy speech observations, an adaptive band selection criterion is developed. Next, on each resulting band-limited noisy speech signal, a repeated autocorrelation operation is carried out, which not only reduces the effect of noise but also strengthens the dominant poles corresponding to the formant frequencies. Finally, spectrum of the RACF is computed and instead of direct spectral peak picking, a model fitting scheme is introduced to find out model parameters which lead to formant estimation. The proposed algorithm has been tested on natural vowels as well as some naturally spoken sentences in the presence of different environmental noises. It is found that the proposed scheme provides better formant estimation accuracy in comparison to some of the existing methods at low levels of signal-to-noise ratio.

Most strain imaging techniques follow a pipeline strategy: in the first step, tissue displacement is estimated from radio-frequency (RF) frames, and in the second step, a spatial derivative operation is applied. There are two main issues that arise from this framework. First, the gradient operation amplifies noise, and therefore, smoothing techniques have to be adopted. Second, strain estimation does not exploit the original RF data. It rather relies solely on the noisy displacement field. In this paper, a novel technique is proposed that utilizes both the displacement field and the RF frames to accurately obtain the strain estimates. The normalized cross correlation (NCC) metric between two corresponding windows around the samples of the pre- and post-compressed images is employed to generate a dissimilarity measurement. The derivative of NCC with respect to the strain is analytically derived using the chain rule. This allows an efficient minimization of the dissimilarity metric with respect to the strain using the gradient descent optimization technique. The effectiveness of the proposed method is investigated through simulation data, phantom experiments, and in vivo patient data. The experimental results show that exploiting the information in RF data significantly improves the strain estimates.

Salient regions are the most dominant parts of an image, which capture human visual system's attention. Finding computational methods that are able to detect salient regions especially in images with messy background is a challenging task. In this paper, a novel segment-based saliency detection method using the wavelet transform is proposed. The human beings are attracted by objects or regions rather than individual pixels. Moreover, at pixel grid, a sudden change in a pixel from a cluttered scene obtains a high saliency value, whereas at segment grid, the saliency is determined by considering each pixel and its neighboring pixels. Thus, applying fast and efficient frequency transforms at the segment grid can improve the capability of the method in images with cluttered background or repeating distractors. The proposed method is evaluated on several images from a publicly available dataset of natural images. Experimental results show that the proposed method provides larger values of area under the receiver operating characteristic curve, precision-recall, and F-measure in comparison to some of the state-of-the-art methods.

Single-microphone speech dereverberation is a challenging problem of de-convolving the reverberation produced by the room impulse response from the speech signal, when only one observation of the reverberant signal (one microphone) is available. By using linear prediction (LP)-residuals and spectral subtraction as two promising tools for dereverberation, a new technique is proposed. The first stage of the proposed technique consists of pre-whitening followed by a delayed long-term LP filtering whose kurtosis or skewness of LP-residuals is maximized to control the weight updates of the inverse filter. A nonlinear spectral subtraction scheme is the second stage of the proposed technique. It is shown that that the proposed algorithms outperform the existing major single-microphone methods in terms of a number of qualitative and quantitative measures.

In synthetic aperture radar (SAR) images, speckle noise reduction is a crucial pre-processing step for their successful interpretation and thus has drawn a great deal of attention of researchers in the image processing community. The Bayesian estimation is a powerful signal estimation technique and has been widely used for speckle noise removal in images. In this work, a low complexity wavelet-based Bayesian estimation technique for despeckling of images is developed. The main idea of the proposed technique is in establishing suitable statistical models for the wavelet coefficients and then in using these models to develop a shrinkage function with a low-complexity realization for the estimation of the wavelet coefficients of the noise-free images. The experimental results demonstrate the effectiveness of the proposed despeckling scheme in providing a significant reduction in the speckle noise at a very low computational cost and simultaneously preserving the image details.

Dissimilarities in equal error rates (EERs) of multiple matchers heavily influence the performance of multi-biometric systems. A normalization technique aims at improving the recognition rate of such a system. In view of this, in this paper, an anchored normalization technique, referred to as improved anchored min-max (IAMM) technique for a multimodal biometric system, is developed. In the proposed technique, the anchor value is computed from the raw matching score sets corresponding to each of the modalities used in the system. This anchor value does not require a priori knowledge of the equal error rates and genuine/impostor score distributions of the individual matchers used in the system. It takes into account the average and variations of the score values that occur more than once in each score set. The performance of IAMM, in terms of EER and genuine acceptance rates @10% and @20% false acceptance rates, is evaluated on a multi-biometric system. The experimental results show that the performance of a multi-biometric system using the proposed normalization technique is superior to that of the uni-biometric systems or to that of the system using the existing normalization techniques.

In joint source-channel arithmetic coding, a forbidden symbol (FS) is added to the symbol set to make it more robust against transmission errors. By splitting the interval occupied by the FS into subintervals, various configurations are possible. In this paper, the delay probability function (DPF), the probability of the number of bits required to detect an error, is calculated for various FS configurations. A figure of merit is also proposed for optimizing the FS configuration. It determines the probability of missed error detection (PMD). Simulations are carried out by employing a breadth-first suboptimal sequential MAP. The simulation results show the effectiveness of the proposed figure of merit, and support the FS configuration in which the FS interval lies entirely between the other information carrying symbols to be the best.

The success of sparse representation, in face recognition and visual tracking, has attracted much attention in computer vision in spite of its computational complexity. These sparse representation-based methods assume that the coding residual follows either Gaussian or Laplacian distribution, which may not be accurate enough to describe the coding residuals in real scenarios. In order to deal with such issues in visual tracking, a novel generative tracker is proposed in a Bayesian inference framework by exploiting both the robust sparse coding and the principle component analysis (PCA) algorithm. In contrast to the existing algorithms, the proposed method introduces weighted least squares into the PCA reconstruction avoiding the much complex l1-regularization. Further, it is proposed to generate an occlusion map based on weights, and is used to avoid updating the occlusion information during incremental subspace learning. The performance evaluation on the challenging image sequences demonstrates that the proposed method performs favorably when compared with the several state-of-the-art methods.

Feature extraction from each scale of an image pyramid to construct a feature pyramid is considered as a computational bottleneck for many object detectors. In this paper, we present a novel technique for the approximation of feature pyramids in the 2D discrete cosine transform (2DDCT) domain. The proposed method is based on a feature resampling technique in the 2DDCT domain, and exploits the effect of resampling an image on the feature responses. Experimental results show that the proposed scheme provides feature approximation accuracy higher than that of the spatial domain counterpart using gradient magnitude or gradient histograms. Further, when the proposed method is employed for pedestrian detection, it provides a logaverage miss-rate lower than that provided by the state-of-the-art techniques on INRIA, ETH, and TUD datasets and performs favorably on Caltech dataset, while performing in real-time.

Speckle noise reduction is a prerequisite task in images captured by ultrasonography systems due to their inherent noisy nature. In this work, we propose a new despeckling method in the contourlet domain using the Cauchy prior. The multiplicative speckle noise is first transferred to an additive one using a logarithmic transform. The logarithmically-transformed contourlet coefficients of the image and noise are assumed to be the Cauchy and Maxwell distributions, respectively. In order to estimate the noise-free contourlet coefficients, an efficient closed-form Bayesian maximum a posteriori estimator is developed. Simulations are carried out to evaluate the performance of the proposed despeckling method by using the synthetically-speckled and real ultrasound images. It is shown that the proposed method outperforms several existing techniques in terms of the signal-to-noise ratio and is able to preserve the diagnostically signific ant details of the ultrasound images.

In a distributed video coding problem, use of a correct model for the correlation noise plays a significant role in improving the decoding performance and consequently in providing higher coding efficiency. In this work, we first study the predictive and additive correlation noise models at the DCT coefficient band level for transform-domain distributed video coding. Then, bounds on compression rates for encoding the quantized DCT coefficient band are obtained for both the correlation noise models. We then investigate how the distribution of the DCT coefficient bands in each WZ frame affects the compression rate bound in each correlation noise model. It is shown that for the DCT coefficient bands being non-uniformly distributed, the compression rate bound in the additive correlation noise model lower than that in the predictive one. Moreover, it is shown that selecting a wrong correlation model leads to compression rate loss in the decoder. The simulation results are provided to validate the theoretical investigation.

The performance of multiple matchers heavily influence the recognition accuracy of a multi-biometric system under the simple sum-rule-based score-level fusion. In this paper, a weighting technique, referred to as score reliability based weighting (SRBW) technique, is developed to estimate weights for the matchers in order to improve the recognition rate of multi-biometric systems at the score level. In the proposed technique, the reliabilities are computed directly from the raw matching scores obtained from the individual matchers. The proposed weighting technique does not require a priori knowledge of the rankings of matching scores, or the equal error rates, or the genuine/impostor score distributions of the individual matchers used in the system. The experimental results show that the performance of a multi-biometric system using the proposed weighting technique is superior to that of the uni-biometric systems or to that of the multi-biometric systems using the existing weighting techniques in terms of equal error rate and genuine acceptance rate at 1% false acceptance rate.

In recent years, many works on digital image watermarking have been proposed all aiming at protection of the copyright of an image document or authentication of data. This paper proposes a novel watermark decoder in the contourlet domain . It is known that the contourlet coefficients of an image are highly non-Gaussian and a proper distribution to model the statistics of the contourlet coefficients is a heavy-tailed PDF. It has been shown in the literature that the normal inverse Gaussian (NIG) distribution can suitably fit the empirical distribution. In view of this, statistical methods for watermark extraction are proposed by exploiting the NIG as a prior for the contourlet coefficients of images. The proposed watermark extraction approach is developed using the maximum likelihood method based on the NIG distribution. Closed-form expressions are obtained for extracting the watermark bits in both clean and noisy environments. Experiments are performed to verify the robustness of the proposed decoder. The results show that the proposed decoder is superior to other decoders in terms of providing a lower bit error rate. It is also shown that the proposed decoder is highly robust against various kinds of attacks such as noise, rotation, cropping, filtering, and compression.

There are a number of image denoising methods in the wavelet domain using statistical models. It is known that the performance of such methods can be significantly improved by taking into account the statistical dependencies between the wavelet coefficients. It is shown that the vector-based hidden Markov model (VB-HMM) is capable of capturing both the subband marginal distribution and the inter-scale, intra-scale and cross orientation dependencies of the wavelet coefficients. In view of this, we propose a new maximum a posteriori estimator using the VB-HMM as a prior for the wavelet coefficients of images. This is realized by deriving an efficient closed-form expression for the shrinkage function. Experimental results are performed to evaluate the performance of the proposed denoising method. The results demonstrate that the proposed method outperforms some of the state-of-the-art techniques in terms of both the peak signal to noise ratio and perceptual quality.

The success of sparse representation in face recognition has motivated the development of sparse representation-based appearance models for visual tracking. These sparse representation-based trackers show state-of-the-art performance, but at the cost of computationally expensive l1-norm minimization. As the computational cost prevents the tracker from being used in real-time systems such as real-time surveillance and military operations, it has become a very important issue. With the aim of reducing the computational complexity of l1-norm minimization, a structural local DCT sparse appearance model is proposed in a particle filter framework. Application of DCT on local patches helps to reduce the dimensions of the dictionary as well as candidate samples by using low-pass filtered DCT coefficients. This in turn helps to remove the information relating to occlusion and background clutter thereby reducing the ambiguity created while computing the confidences of the target samples. The proposed method is evaluated on the challenging image sequences available in the literature and its performance compared with three recent state-of-the-art methods. It is shown that the proposed method provides superior/similar performance for most of the sequences with reduced computational complexity in l1-norm minimization.

Speckle reduction has been a prerequisite for many SAR image processing tasks. This work presents a new approach for despeckling of SAR images in the contourlet domain using the alpha-stable distribution. It is shown that the alpha-stable distribution provides a good fit for the contourlet coefficients of an image, since it can capture the large peak and heavy tails of the distribution of the empirical data. This model is then exploited in a Bayesian maximum a posteriori estimator to restore the noise-free contourlet coefficients. The performance of the proposed despeckling method is evaluated using synthetically-speckled and real SAR images. Simulations are carried out using synthetically speckled images to investigate the performance of the proposed method, and compare it with that of some of the existing methods. The experimental results show that the proposed method can provide better preservation of the edges and can yield better visual quality as compared to some of the existing methods.

Digital watermarking has been widely used in the copyright protected images in multimedia. This paper addresses the blind watermark detection problem in contourlet domain. It is known that the contourlet coefficients of images have non-Gaussian property and can be well modelled by non-Gaussian distributions such as the normal inverse Gaussian (NIG). In view of this, we exploit this model to derive closed-form expressions for the test statistics and design an optimum blind watermark detector in the contourlet domain. Through conducting several experiments, the performance of the proposed detector is evaluated in terms of the probabilities of detection and false alarm and compared to that of the other existing detectors. It is shown that the proposed detector using the NIG distribution is superior to other detectors in terms of providing higher rate of detection. It is also shown that the proposed NIG-based detector is more robust than other detectors against attacks, such as JPEG compression and Gaussian noise.

A new contourlet-based method is introduced for reducing noise in images corrupted by additive white Gaussian noise. This method takes into account the statistical dependencies among the contourlet coefficients of different scales. In view of this, a non-Gaussian multivariate distribution is proposed to capture the across-scale dependencies of the contourlet coefficients. This model is then exploited in a Bayesian maximum a posteriori estimator to restore the clean coefficients by deriving an efficient closed-form shrinkage function. Experimental results are performed to evaluate the performance of the proposed denoising method using typical noise-free images contaminated by simulated noise. The results show that the proposed method outperforms some of the state-of-the-art methods in terms of both the subjective and objective criteria.

In his paper a method of seizure detection has been proposed based on the Discrete Wavelet Transform (DWT) analysis of the dominant Intrinsic mode function(IMF) resulting from the Empirical Mode Decomposition(EMD) of the EEG signals. Considering the normalized energy, Fourier spectrum and cross-correlation coefficient analysis, only the 4th Level DWT coefficients of the dominant IMF is found reasonable for feature computation. In order to reduce the dimension of the feature vector, Higher order statistics of these coefficients are employed to form he feature vector. The reduced feature vector thus formed is found effective for distinguishing seizure and non-seizure EEG signals when fed to a k-nearest neighborhood (k-NN) classifier. Extensive simulations are carried out using a benchmark EEG dataset. It is shown that the proposed method is capable of producing greater sensitivity, specificity and accuracy in comparison to that obtained by a sate-of-the-art method using the same EEG dataset and classifier.

MOTIVATION: Insertion/deletion (indel) and amino acid substitution are two common events that lead to the evolution of and variations in protein sequences. Further, many of the human diseases and functional divergence between homologous proteins are more related to indel mutations, even though they occur less often than the substitution mutations do. A reliable identification of indels and their flanking regions is a major challenge in research related to protein evolution, structures and functions. RESULTS: In this article, we propose a novel scheme to predict indel flanking regions in a protein sequence for a given protein fold, based on a variable-order Markov model. The proposed indel flanking region (IndelFR) predictors are designed based on prediction by partial match (PPM) and probabilistic suffix tree (PST), which are referred to as the PPM IndelFR and PST IndelFR predictors, respectively. The overall performance evaluation results show that the proposed predictors are able to predict IndelFRs in the protein sequences with a high accuracy and F1 measure. In addition, the results show that if one is interested only in predicting IndelFRs in protein sequences, it would be preferable to use the proposed predictors instead of HMMER 3.0 in view of the substantially superior performance of the former.

BACKGROUND: The alignment of multiple protein sequences is one of the most commonly performed tasks in bioinformatics. In spite of considerable research and efforts that have been recently deployed for improving the performance of multiple sequence alignment (MSA) algorithms, finding a highly accurate alignment between multiple protein sequences is still a challenging problem. RESULTS: We propose a novel and efficient algorithm called, MSAIndelFR, for multiple sequence alignment using the information on the predicted locations of IndelFRs and the computed average log-loss values obtained from IndelFR predictors, each of which is designed for a different protein fold. We demonstrate that the introduction of a new variable gap penalty function based on the predicted locations of the IndelFRs and the computed average log-loss values into the proposed algorithm substantially improves the protein alignment accuracy. This is illustrated by evaluating the performance of the algorithm in aligning sequences belonging to the protein folds for which the IndelFR predictors already exist and by using the reference alignments of the four popular benchmarks, BAliBASE 3.0, OXBENCH, PREFAB 4.0, and SABRE (SABmark 1.65). CONCLUSIONS: We have proposed a novel and efficient algorithm, the MSAIndelFR algorithm, for multiple protein sequence alignment incorporating a new variable gap penalty function. It is shown that the performance of the proposed algorithm is superior to that of the most-widely used alignment algorithms, Clustal W2, Clustal Omega, Kalign2, MSAProbs, MAFFT, MUSCLE, ProbCons and Probalign, in terms of both the sum-of-pairs and total column metrics.

An improved fast iterative shrinkage thresholding algorithm (IFISTA) for image deblurring is proposed. The IFISTA algorithm uses a positive definite weighting matrix in the gradient function of the minimization problem of the known fast iterative shrinkage thresholding (FISTA) image restoration algorithm. A convergence analysis of the IFISTA algorithm shows that due to the weighting matrix, the IFISTA algorithm has an improved convergence rate and improved restoration capability of the unknown image over that of the FISTA algorithm. The weighting matrix is predetermined and fixed, and hence, like the FISTA algorithm, the IFISTA algorithm requires only one matrix vector product operation in each iteration. As a result, the computational burden per iteration of the IFISTA algorithm remains the same as in the FISTA algorithm. Numerical examples are presented that demonstrate the improved performance of the IFISTA algorithm over that of the FISTA and iterative shrinkage thresholding (ISTA) algorithms in terms of the convergence speed and the peak signal-to-noise ratio.

This paper presents a speech enhancement approach, where an adaptive threshold is statistically determined based on Student$t$Modeling of Teager energy (TE) operated perceptual wavelet packet (PWP) coefficients of noisy speech. In order to obtain an enhanced speech, the threshold thus derived is applied upon the PWP coefficients by employing a Student$t$pdf dependent custom thresholding function, which is designed based on a combination of modified hard and semisoft thresholding functions. Extensive simulations are carried out using the NOIZEUS database to evaluate the effectiveness of the proposed method for car and multi-talker babble noise corrupted speech signals. Several standard objective measures and subjective evaluations including formal listening tests show that the proposed method outperforms some of the state-of-the-art speech enhancement methods at high as well as low levels of SNRs.

The past decade has witnessed significant progress in object detection and tracking in videos. In this paper, we present a model for collaboration between a pre-trained object detector and multiple single object trackers in the particle filter tracking framework. For each frame, we construct an association between the trackers and the detections, and when a tracker is successfully associated to a detection, we treat this detection as the key-sample for this tracker. We present a dual motion model that incorporates the associated detections with the object dynamics. Then, a likelihood function provides different weights for the propagated and the newly created particles, reducing the effect of false positives and missed detections in the tracking process. In addition, we use generative and discriminative appearance models to maximize the appearance variation among the targets. The performance of the proposed algorithm compares favorably with that of the state-of-the-art approaches on three public sequences.

The wavelet coefficients of images show heavy-tailed marginal statistics as well as strong inter- and intra-subbands and across orientations dependencies. The vector-based hidden Markov model (HMM) has been shown to be an effective statistical model for wavelet coefficients, which is capable of capturing both the subband marginal distribution and the inter-scale and intra-scale dependencies of the wavelet coefficients. In this paper, we propose a locally-optimum watermark detector using the HMM model for image wavelet coefficients. The performance of the proposed detector is studied through simulation and is shown to be superior to that of other detectors in terms of the imperceptibility of the embedded watermark and detection rate.

A space time code based partial rank affine projection (PRAP) algorithm is proposed. The proposed algorithm uses an input signal where the input signal matrix Xkbecomes an orthogonal matrix. For this input signal, matrix (XkTXk) becomes a diagonal matrix whose inverse can be easily computed. Thus, the proposed algorithm saves a significant amount of computations. Due to this feature the proposed PRAP algorithm is shown to offer a faster convergence speed and a smaller computational burden per iteration than the NLMS algorithm does.

It is known that the contourlet coefficients of images have non-Gaussian property and heavy tails. In view of this, an appropriate distribution to model the statistics of the contourlet coefficients would be the one having large peaks, and tails heavier than that of a Gaussian PDF, i.e., a heavy-tailed PDF. This paper proposes a new image modeling in the contourlet domain, where the magnitudes of the coefficients are modeled by a symmetric alpha-stable distribution which is best suited for modeling transform coefficients with a high non-Gaussian property and heavy tails. It is shown that the alpha-stable family of distributions provides a more accurate model to the contourlet subband coefficients than the formerly used distributions, namely, the generalized Gaussian and Laplacian distributions, both in terms of the subjective measure of the Kolmogorov-Smirnov distance and the objective measure of comparing the log-scale histograms.

This paper presents a speaker-independent speech emotion recognition method, where emotional features are derived from the Teager energy (TE) operated wavelet coefficients of speech signal. Due to TE operation, the enhanced detail as well as approximate Wavelet coefficients thus obtained is then used to compute entropy. Entropy values of TE operated detail and approximate wavelet coefficients not only reduces feature dimension but also form an effective feature vector for distinguishing different emotions when fed to a Euclidean distance based classifier. Extensive simulations are carried out using EMO-DB German speech emotion database containing four class emotions, such as angry, happy, sad and neutral. Simulation results show that the proposed method is capable of outperforming an existing speaker-independent emotion recognition method thus solving a four-class emotion recognition problem in terms of higher recognition accuracy with lower computation.

In the past decade, several schemes for digital image watermarking have been proposed to protect the copyright of an image document or to provide proof of ownership in some identifiable fashion. This paper proposes a novel multiplicative watermarking scheme in the contourlet domain. The effectiveness of a watermark detector depends highly on the modeling of the transform-domain coefficients. In view of this, we first investigate the modeling of the contourlet coefficients by the alpha-stable distributions. It is shown that the univariate alpha-stable distribution fits the empirical data more accurately than the formerly used distributions, such as the generalized Gaussian and Laplacian, do. We also show that the bivariate alpha-stable distribution can capture the across scale dependencies of the contourlet coefficients. Motivated by the modeling results, a blind watermark detector in the contourlet domain is designed by using the univariate and bivariate alpha-stable distributions. It is shown that the detectors based on both of these distributions provide higher detection rates than that based on the generalized Gaussian distribution does. However, a watermark detector designed based on the alpha-stable distribution with a value of its parameter α other than 1 or 2 is computationally expensive because of the lack of a closed-form expression for the distribution in this case. Therefore, a watermark detector is designed based on the bivariate Cauchy member of the alpha-stable family for which α = 1 . The resulting design yields a significantly reduced-complexity detector and provides a performance that is much superior to that of the GG detector and very close to that of the detector corresponding to the best-fit alpha-stable distribution. The robustness of the proposed bivariate Cauchy detector against various kinds of attacks, such as noise, filtering, and compression, is studied and shown to be superior to that of the generalized Gaussian detector.

In this paper, a novel involutory parametric transform is proposed by exploiting the reciprocal-orthogonal parametric transform. In addition, a recursive algorithm is proposed for its simple construction and fast computation. The transform has a very large number of independent parameters that are useful for many applications. Specifically, we show by implementing the double random phase encoding technique that the independent parameters of the proposed transform can successfully be used as an additional secret key for image encryption.

Amyotrophic lateral sclerosis (ALS) is a common fatal motor neuron disease that assails the nerve cells in the brain. As the nervous system controls the muscle activity, the electromyography (EMG) signals can be viewed and examined in order to detect the vital features of the ALS disease in individuals. In this paper, the discrete wavelet transform (DWT) based features, which are extracted from a frame of EMG data, are introduced to classify the normal person and the ALS patients. From each frame of EMG data, instead of using a large number of DWT coefficients, the DWT coefficients with higher values as well as their mean and maxima are proposed to be used, which drastically reduces the feature dimension. It is shown that the proposed feature vector offers a high within class compactness and between class separations. For the purpose of classification, the K-nearest neighborhood classifier is employed. In order to demonstrate the classification performance, an EMG database consisted of 5 normal subjects and 5 ALS patients is considered and it is found that the proposed method is capable of distinctly separating the ALS patients from the normal persons.

In this paper, a phase spectrum based acoustic parameter is presented for the detection of nasalized vowels from the mixture of oral and nasalized vowels of normal speakers. Acoustic analysis shows that during the event of nasalization, although additional formants (resonances) at various frequency locations are introduced, the introduction of a new formant in low frequency region around 250 Hz is found to remain consistent irrespective of female or male speakers in the modified group delay derived from the phase spectrum. By exploiting and verifying this fact on the band-limited modified group delay spectrum capable of resolving two closely spaced formants, an acoustic parameter RMGD is derived. Utilizing RMGD, the problem of detecting nasalized vowels is solved based on a threshold based scheme or a Euclidean distance based classifier. Simulation Results on TIMIT database show that the proposed method even with a simple classifier is superior in performance in comparison to that of the methods using Mel-frequency cepstral coefficients as a feature and Hidden Markov Modeling or Support Vector Machine as a classifier.

Regulatory interactions among genes and gene products are dynamic processes and hence modeling these processes is of great interest. Since genes work in a cascade of networks, reconstruction of gene regulatory network (GRN) is a crucial process for a thorough understanding of the underlying biological interactions. We present here an approach based on pairwise correlations and lasso to infer the GRN, taking into account the variable time delays between various genes. The proposed method is applied to both synthetic and real data sets, and the results on synthetic data show that the proposed approach outperforms the current methods. Further, the results using real data are more consistent with the existing knowledge concerning the possible gene interactions.

The number of projections is a critical factor in tomographic imaging. The larger the number, the better the quality of the reconstructed image; however, it increases the radiation dose delivered to the patient. Therefore, it is important to keep the number of projections as small as possible. Traditionally, the projections are taken by moving the x-ray source around the patient at uniform angular steps. Taking projections at nonuniform steps may result in better images as compared with that obtained using uniform projections. This paper describes two different approaches that adjust the step size to adaptively select the angle of projections. The first one is based on the spectral richness of the acquired projections and the second relies on the amount of new information added by successive projections. The superior performance of the two proposed methods over the uniform projection scheme is demonstrated through simulation results using both phantom and real images.

This paper presents an approach to detect voice disorders based on wavelet and prosody-related voice properties. First, several statistical measures of the normalized energy contents of the Discrete Wavelet Transform (DWT) coefficients over all voice frames are determined. Then, similar statistical measures of some prosody-related voice properties, such as mean pitch, jitter and shimmer are also computed over all the frames. In order to form a feature vector to be used in both training and testing phases, a set of statistical measure of the normalized energy contents of the DWT coefficients is combined with a set of statistical measure of the extracted prosody-related voice properties. Here, the voice samples under consideration are assumed to be of two categories, namely healthy and disordered thus formulating the problem in the proposed method as a two-class problem to be solved. Finally, the feature vector as obtained above is fed to an Euclidean Distance based classifier to detect the disordered voice. By performing extensive simulations, it is shown that the statistical analysis based on wavelet and prosody-related properties are able to provide effective detection of a variety of voice disorders from the mixture of healthy and disordered voices.

In this paper, a method for the estimation of pitch from noise-corrupted speech observations based on extracting a pitch harmonic and the corresponding harmonic number is proposed. Starting from the harmonic representation of clean speech, a simple yet accurate harmonic sinusoidal autocorrelation (HSAC) model is first derived. By employing this HSAC model expressed in terms of the pitch harmonics of the clean speech, a new autocorrelation-domain least-squares fitting optimization technique is developed to extract a pitch harmonic from the noisy speech. Then, the harmonic number associated with the pitch harmonic is determined by maximizing an objective function formulated as an impulse-train weighted symmetric average magnitude sum function (SAMSF) of the noisy speech. The period of the impulse-train is governed by the estimated pitch harmonic and the maximization of the objective function is carried out through a time-domain matching of periodicity of the impulse-train with that of the SAMSF. An SAMSF-based pitch tracking scheme using dynamic programming is devised to obtain a smoothed pitch contour. In order to demonstrate the efficacy of the proposed method, simulations are conducted by considering naturally spoken speech signals in the presence of white or multi-talker babble noise at different signal-to-noise ratio (SNR) levels. A comprehensive evaluation of the pitch estimation results shows the superiority of the proposed method over some of the state-of-the-art methods under low levels of SNR.

The regulation of gene expression is a dynamic process, hence it is of vital interest to identify and characterize changes in gene expression over time. We present here a general statistical method for detecting changes in microarray expression over time within a single biological group and is based on repeated measures (RM) ANOVA. In this method, unlike the classical F-statistic, statistical significance is determined taking into account the time dependency of the microarray data. A correction factor for this RM F-statistic is introduced leading to a higher sensitivity as well as high specificity. We investigate the two approaches that exist in the literature for calculating the p-values using resampling techniques of gene-wise p-values and pooled p-values. It is shown that the pooled p-values method compared to the method of the gene-wise p-values is more powerful, and computationally less expensive, and hence is applied along with the introduced correction factor to various synthetic data sets and a real data set. These results show that the proposed technique outperforms the current methods. The real data set results are consistent with the existing knowledge concerning the presence of the genes. The algorithms presented are implemented in R and are freely available upon request.

In this paper, first, a new Laplacian kernel is developed to integrate into it the anisotropic behavior to control the process of forward diffusion in horizontal and vertical directions. It is shown that, although the new kernel reduces the process of edge distortion, it nonetheless produces artifacts in the processed image. After examining the source of this problem, an analytical scheme is devised to obtain a spatially varying kernel that adapts itself to the diffusivity function. The proposed spatially varying Laplacian kernel is then used in various nonlinear diffusion filters starting from the classical Perona-Malik filter to the more recent ones. The effectiveness of the new kernel in terms of quantitative and qualitative measures is demonstrated by applying it to noisy images.

In this paper, the problem of joint space-time parameter estimation for underwater wireless communication channels in a multipath environment is addressed. We consider the receive antenna array to be configured with multiple vector sensors, each of which consists of a pair of orthogonal velocity sensors. A quadrilinear model for the channel is formulated, and a quadrilinear decomposition method developed for joint angle and delay estimation (JADE). In addition, a computationally simple subspace-based algorithm is proposed for the problem under consideration. The basic idea behind this algorithm is to use the angle information embedded in the velocity vector sensors to start two polynomial rooting procedures for the angle and delay in succession. Simulation results are finally presented to verify the efficacy of the proposed algorithms.

In this paper, an efficient algorithm for fast computation of the conjugate symmetric sequency-ordered complex Hadamard transform (CS-SCHT) of any length that is a power of two is proposed using the Kronecker product. Since the CS-SCHT matrix is factored into a product of sparse matrices, the resulting structure for the algorithm is very attractive for implementation and similar to that of the well-known Walsh-Hadamard transform, except for some multiplications by -1 or (-√(-1)). It is shown that the proposed N-point complex-valued CS-SCHT algorithm requires Nlog2(N) complex additions/subtractions and (N/2-1) multiplications by (-√(-1)).

In this paper, a one-parameter eight-point orthogonal transform suitable for image compression is proposed. An algorithm for its fast computation is developed and an efficient structure for a simple implementation valid for all possible values of its independent parameter is proposed. It is shown that an appropriate selection of the values of the parameter results in a number of new multiplication-free transforms having a good compromise between the computational complexity and performance. Applying the proposed transform to image compression, we show that it outperforms the existing transforms having complexities similar to that of the proposed one.

This letter investigates the problem of joint estimation of the direction of departure (DOD) and the direction of arrival (DOA) for multi-input multi-output (MIMO) array systems. A new bistatic MIMO array system, configured with multiple transmit sensors and multiple velocity receive sensors is introduced, and a new joint DOD and DOA estimation algorithm is proposed. The key idea behind the proposed algorithm is to use the DOA information embedded in the velocity sensors to start 1-D MUSIC searches for the DOD and DOA in succession. The proposed successive MUSIC algorithm is suitable for irregular array geometry, imposes less constraint on the receive sensor spacing, and requires no parameter pairing nor two-dimensional searching.

Discrete transforms have been widely used in various signal processing applications. Specifically, transform-based data encryption techniques have become attractive for many recent communication systems. In this paper, we propose a fast and efficient image encryption method based on the reciprocal-orthogonal parametric (ROP) transform. By exploiting the properties of the ROP transform, we show that its independent parameters can successfully be used as an additional secret key for encryption. Experiments results carried clearly show the efficiency of the proposed encryption method.

It is well established that the function of human gene can be identified by working on the corresponding gene in a model organism. Such comparative genomic studies have provided new insights into human biology and gene expression. Due to the explosion of genomic data in recent times, highly effective computational comparative genomic algorithms are in greater demand. In this research, a digital signal processing approach using statistically optimal null filter (SONF) is developed for comparative genomic analysis. The instantaneous matched filter in SONF determines the degree of local alignment between the genomic sequences being compared. Through examples the effectiveness of the proposed approach is illustrated in comparison with the other existing convolution based method. In particular, the proposed method is highly efficient in locating a short motif in a large genomic sequence.

A novel framework of the motion-compensated 3-D wavelet transform (MC3DWT) for video denoising is presented in this paper. The motion-compensated temporal wavelet transform is first performed on a sliding window of video frames consisting of previously denoised frames and the current noisy frame. The 2-D spatial wavelet transform is then performed on the temporal subband frames, thus realizing a 3-D wavelet transform. Any of established wavelet-based still image denoising algorithms can then be applied to the high-pass 3-D subbands. The operation of the inverse 2-D spatial wavelet transform followed by the inverse temporal wavelet transform reconstructs the video frames in the buffer. The denoised current frame may be used as an output for real-time processing; meanwhile, the past frames can be updated, one of which may be used as a delayed output for post-processing or for real-time processing that allows some amount of delay. The proposed MC3DWT framework integrates both the spatial filtering and recursive temporal filtering into the 3-D wavelet domain and effectively exploits both the spatial and temporal redundancies. Experimental results have demonstrated a superior visual and quantitative performance of the proposed scheme for various levels of noise and motion.

Generalized singular value decomposition (GSVD) has been used for linear discriminant analysis (LDA) to solve the small sample size problem in pattern recognition. However, this algorithm may suffer from the over-fitting problem. In this paper, we propose a novel orthogonalization technique for the LDA/GSVD algorithm to address the over-fitting problem. In this technique, an orthogonalization of the basis of the discriminant subspace derived from the LDA/GSVD algorithm is carried out through an eigen-decomposition of a small size inner product matrix. It is computationally efficient when data are high dimensional. The technique is further applied to the kernelized LDA/GSVD algorithm, mGSVD-KDA, leading to a new algorithm, referred to as GSVD-OKDA. It is shown that with linear and nonlinear kernels, this new algorithm successfully overcomes the over-fitting problem of the LDA/GSVD and mGSVD-KDA algorithms. Simulation results show that the proposed algorithms provide high recognition accuracy with low computational complexity.

Formant frequency is a one of the most important speech feature, which has widespread applications in speech recognition, synthesis, and compression. In this paper, a new time-frequency domain scheme for the estimation of formant frequencies from noise-corrupted speech signals is presented. In order to overcome the adverse effect of noise, instead of conventional autocorrelation function (ACF), a repeated ACF (RACF) of the noisy speech is employed. Exploiting the characteristics of the zero lag, a set of equations containing the lower lags of the RACF of the noisy speech is used to estimate the formant frequencies. In order to avoid estimation errors that may occur in the case of weak formants, a frequency-domain algorithm is introduced utilizing the RACF of the observed speech. Formant frequency estimation accuracy is measured for different natural and synthetic vowels in noisy environments and even at low levels of signal-to-noise ratio, a better performance is obtained by the proposed scheme in comparison to some of the existing methods.

A new bivariate maximum a posteriori estimator is proposed for the magnitude components of the dual-tree complex wavelet transform (DT-CWT) coefficients in order to reduce additive white Gaussian noise in a video. The estimator considers the fact that the magnitude components of the DT-CWT coefficients of the Gaussian distributed noise fit the generalized Gamma distribution very well. For spatial filtering, the joint distribution function of the magnitude components of the DT-CWT coefficients of the two neighboring frames of a video is considered to be locally-adaptive bivariate Gaussian having a non-negative mean. The correlation coefficient of this distribution function acts as an indirect measure of the motion of the DT-CWT coefficients between two neighboring frames. A recursive time averaging of the spatially filtered magnitude components is adopted for further noise reduction. Experimental results on test video sequences show that the proposed estimator provides an average peak signal-to-noise ratio that is higher than that provided by the others.

An effective method for pitch estimation from severely noise-corrupted speech is presented in this paper. We argue that an accurate estimate of a pitch-harmonic (PH) can be obtained from the Discrete Cosine Transform (DCT) power spectrum of an enhanced frame. By exploiting the PH thus extracted, a spectral matching method is developed in the DCT domain to determine the true harmonic number associated to the PH that leads to a pitch estimate. In a heavy noisy scenario, the superior efficacy of the proposed pitch estimation method relative to some of the existing ones is confirmed through simulation results using theKeeledatabase.

Generalized singular value decomposition (GSVD) has been used in the literature for linear discriminant analysis (LDA) to solve the small sample size problem in pattern recognition. However, this algorithm suffers from excessive computational load when the sample dimension is high. In this paper, we present a modified version of the LDA/GSVD algorithm to enhance the computational efficiency, referred to as EGSVD-LDA algorithm, which uses the linear combination of the sample vectors to represent the singular vectors so as to circumvent the calculation of the high dimensional singular vectors through SVD. Further, to overcome the over-fitting problem of the GSVD-based algorithms, we have also proposed a new method to orthogonalize the discriminative subspace derived from the GSVD framework through a Gram-Schmidt process in an inner product space. These methods are efficient when data are high dimensional. Simulation results show that the EGSVD-LDA algorithm, especially its orthogonalized version, overcomes the computational complexity problem and provides high recognition accuracy with low computational load.

Ultrasound imaging is widely used for diagnostic purposes among the clinicians. A major problem concerning the ultrasound images is their inherent corruption by the multiplicative speckle noise that hampers the quality of the diagnosis, and reduces the efficiency of the algorithms for automatic image processing. In this paper, we propose a new spatially adaptive wavelet-based method in order to reduce the speckle noise from ultrasound images. A spatially adaptive threshold is introduced for denoising the coefficients of log-transformed ultrasound images. The threshold is obtained from a Bayesian maximum a posteriori estimator that is developed using a symmetric normal inverse Gaussian probability density function (PDF) as a prior for modelling the coefficients of the log-transformed reflectivity. A simple and fast method is provided to estimate the parameters of the prior PDF from the neighbouring coefficients. Extensive simulations are carried out using synthetically speckled and ultrasound images. It is shown that the proposed method outperforms several existing techniques in terms of the signal-to-noise ratio, edge preservation index and structural similarity index and visual quality, and in addition, is able to maintain the diagnostically significant details of ultrasound images.

Traditional statistical detectors of the discrete wavelet transform (DWT)-based image watermarking use probability density functions (PDFs) that show inadequate matching with the empirical PDF of image coefficients in view of the fact that they use a fixed number of parameters. Hence, the decision values obtained from the estimated thresholds of these detectors provide substandard detection performance. In this paper, a new detector is proposed for the DWT-based additive image watermarking, wherein a PDF based on the Gauss-Hermite expansion is used, in view of the fact that this PDF provides a better statistical match to the empirical PDF by utilizing an appropriate number of parameters estimated from higher-order moments of the image coefficients. The decision threshold and the receiver operating characteristics are derived for the proposed detector. Experimental results on test images demonstrate that the proposed watermark detector performs better than other standard detectors such as the Gaussian and generalized Gaussian (GG), in terms of the probabilities of detection and false alarm as well as the efficacy. It is also shown that detection performance of the proposed detector is more robust than the competitive GG detector in the case of compression, additive white Gaussian noise, filtering, or geometric attack.

In this paper, a new algorithm for pitch extraction from noisy speech signals based on both temporal and spectral representations is presented. We derive a harmonic sinusoidal correlation (HSC) model of clean speech as a temporal representation. Given only a noisy speech frame, a noise-robust least-squares minimization technique is proposed to acquire the parameters of the HSC model which are directly employed for the accurate estimation of a pitch-harmonic (PH). Exploiting the extracted PH and based on a spectral representation which is an enhanced spectrum in the discrete cosine transform domain, a two-fold criterion is developed in order to achieve the true consecutive number corresponding to PH that is finally adopted for pitch detection in the presence of noise. Simulation results using the Keele pitch extraction reference database manifest that combining the multi cues obtained from the temporal as well as spectral representations, the proposed algorithm is able to achieve a superior efficacy in comparison to some of the existing methods from high to very low signal-to-noise ratio (SNR) levels.

In this paper, the symmetric normal inverse gaussian (SNIG) probability density function (PDF) is proposed as a highly suitable prior for modelling the DCT coefficients of natural images. A new method, based on minimizing the Kullback-Leibler divergence between the proposed prior and the empirical PDF extracted from image data, is proposed to estimate the SNIG parameters. The efficacy of the proposed parameter estimation technique is tested using Monte-Carlo simulations. It is shown that the SNIG PDF is a more effective prior as compared to the generalized Gaussian (GG), α-stable, and Laplacian PDFs for modelling the full-frame DCT coefficients of natural images. For the block-DCT coefficients, the SNIG PDF is shown to be better than the GG and Laplacian PDFs, and comparable to the α-stable one, while incurring much less complexity for parameter estimation.

In this paper, we define a blind-block reciprocal- transpose matrix operator (BBRT). One of the interesting proprieties of the BBRT matrix is that it is very easy to obtain its inverse. We then propose a new blind-block reciprocal parametric transform and show that its matrix operator is a BBRT matrix. The transform has a large number of independent parameters that are useful for many applications and can specifically be used as an additional secret key for encryption and watermarking. It is shown that the proposed transform reduces to some of the existing transforms and also to new classes of parametric transforms having some desirable proprieties.

This paper presents a new algorithm for the parameter estimation of minimum-phase autoregressive moving average (ARMA) systems from noise-corrupted observations. In order to estimate the AR parameters of the ARMA system, based on a repeated autocorrelation function (ACF) of the observed data, a set of zero lag compensated equations has been developed. For the estimation of the MA parameters, first, a noise-subtraction algorithm is proposed to reduce the effect of noise from the ACF of the residual signal which is obtained by filtering the noisy ARMA signal via the estimated AR parameters. The MA parameters are then estimated by using a spectral factorization corresponding to the noise-compensated ACF of the residual signal. Computer simulations are carried out for ARMA systems of different orders under noisy environments and simulation results demonstrate a superior identification performance in terms of estimation accuracy and consistency.

This paper presents a new technique for the minimization of I/O delay in the architectural synthesis of cyclic data flow graphs (DFG) representing DSP algorithms taking into consideration the inter-processor communication delays. In this paper, the question of optimizing the I/O delay without scarifying the iteration period (throughput) with non-negligible inter-processor communication overhead is addressed. The proposed technique operating on the cyclic DFG of a DSP algorithm is designed to evaluate the relative firing times of the nodes by using Floyd-Warshall's longest path algorithm so that the inter-processor communication overhead is taken into consideration to provide an optimized time and processor schedule. Moreover, the proposed scheme is applied to well- know DSP benchmarks and seen that it is efficient in minimizing the I/O delay without scarifying the iteration period.

Prediction of the protein-coding regions (exons) is one of the central issues of DNA sequence analysis. Most of the existing computational methods exploit the period-3 property of the coding-regions to distinguish exons from noncoding regions (introns). However, the current Discrete Fourier Transform (DFT) based methods are inadequate in predicting short exons. In this paper, we present a model-based exon detection approach using statistically optimal null filter. The proposed method employs a model of the period-3 characteristic to maximize signal-to-noise ratio, and least-squares optimization criteria to rapidly detect the presence of exons in the input DNA sequence. Through examples, it is shown that the proposed method is highly effective as compared to the DFT technique, especially in identifying short exons and successive exons separated by short introns.

A new detector using modified Gauss-Hermite (GH) probability density function (PDF) is proposed for the wavelet-domain image watermarking scheme. It is shown that the proposed PDF matches the empirical one of image wavelet coefficients better than other conventional PDFs such as the generalized Gaussian and Bessel K-form. This is because of the fact that the modified GH PDF utilizes an arbitrary number of higher order moments of the wavelet coefficients instead of considering only the first few for the parameter estimation process. The proposed PDF is then used for designing the statistical detector for a wavelet-based image watermarking algorithm. Experimental results on a standard image database show that the proposed detector provides a higher detection probability and lower false alarm than that provided by the others.

A novel algorithm for pitch frequency (PF) estimation from non-stationary noise-corrupted speech observations is presented in this paper based on both spectral pre-processing and temporal representation. A modified power spectral subtraction based de-noising scheme that allows tracking the time-variation of the underlying non-stationary noise is put forward to enhance speech prior to PF estimation. The de-noised speech is then utilized to propose a squared difference function of the Linear Prediction (LP) residual which is expected to reveal more prominent dips at integral multiples of the pitch period compared to that revealed by the LP residual. The dips at different pitch-harmonic locations are added and weighted by a periodicity dependent weighting factor for every possible pitch period thus yielding a weighted and harmonically summed temporal function which is globally minimized to extract the desired PF. Simulation results using the Keele database show the superior efficacy of the proposed method in the presence of a multi-talker babble noise relative to some of the existing methods.

The probability density functions (PDFs) of the wavelet coefficients play a key role in many wavelet-based image processing algorithms, such as denoising. The conventional PDFs usually have a limited number of parameters that are calculated from the first few moments only. Consequently, such PDFs cannot be made to fit very well with the empirical PDF of the wavelet coefficients of an image. As a result, the shrinkage function utilizing any of these density functions provides a substandard denoising performance. In order for the probabilistic model of the image wavelet coefficients to be able to incorporate an appropriate number of parameters that are dependent on the higher order moments, a PDF using a series expansion in terms of the Hermite polynomials that are orthogonal with respect to the standard Gaussian weight function, is introduced. A modification in the series function is introduced so that only a finite number of terms can be used to model the image wavelet coefficients, ensuring at the same time the resulting PDF to be non-negative. It is shown that the proposed PDF matches the empirical one better than some of the standard ones, such as the generalized Gaussian or Bessel K-form PDF. A Bayesian image denoising technique is then proposed, wherein the new PDF is exploited to statistically model the subband as well as the local neighboring image wavelet coefficients. Experimental results on several test images demonstrate that the proposed denoising method, both in the subband-adaptive and locally adaptive conditions, provides a performance better than that of most of the methods that use PDFs with limited number of parameters.

This paper studies an important problem, namely, the joint multipath routing and rate control for multidescription coded (MD-coded) video streaming in wireless ad hoc networks. In addition to selecting a pair of paths to optimize the expected end-to-end video quality, we also explore an optimal packet skipping strategy for the rate control in order to minimize the impact of the skipped packets on the quality of the video. The R-D hint information, consisting of the size of the packets in bits and the importance of the packets for reconstructing the video, is used to characterize the packets in an R-D sense. Since searching for paths to minimize the expected end-to-end video distortion by simultaneously considering the skipped packets prior to the transmission and those dropped/delayed during the transmission is a highly complex problem and is expected to be NP-hard, we develop a heuristic greedy-relaxation-based routing solution that enables the system to efficiently select near-optimal paths. Extensive simulation studies have been conducted to compare the performance of the proposed algorithm with that of several existing algorithms, showing the superior performance of the proposed one. Such a joint rate control and multipath routing approach provides an important methodology for high-quality real-time video streaming applications over ad hoc wireless networks.

A major problem in medical ultrasonography is the inherent corruption of ultrasound images with speckle noise that severely hampers the diagnosis and automatic image processing tasks. In this paper, an efficient wavelet-based method is proposed for despeckling medical ultrasound images. A closed-form Bayesian wavelet-based maximum a posteriori denoiser is developed in a homomorphic framework, based on modelling the wavelet coefficients of the log-transform of the reflectivity with a symmetric normal inverse Gaussian (SNIG) prior. A simple method is presented for obtaining the parameters of the SNIG prior using local neighbors. Thus, the proposed method is spatially adaptive. Experiments are carried out using synthetically speckled and real ultrasound images, and the results show that the proposed method performs better than several other existing methods in terms of the signal-to-noise ratio and visual quality.

A new approach for the formant frequency estimation of the voiced speech segments in the presence of noise is presented in this paper. A correlation model for the voiced speech is proposed considering the vocal-tract system as an autoregressive moving average (ARMA) model with a periodic impulse-train excitation. It is shown that the formant frequencies can be directly obtained from the model parameters. An adaptive residue-based least-squares optimization algorithm is proposed to estimate the model parameters, which overcomes the failure of conventional correlation based techniques in estimating formant frequencies at a low signal-to-noise ratio (SNR). The proposed algorithm has been tested on synthetic and natural vowels as well as voiced segments of some naturally spoken sentences from TIMIT database in presence of white Gaussian or babble noises. The experimental results show that the proposed method is more robust to noise than some existing methods even at a low SNR of 0 dB.

In this paper, we present a robust pitch estimation algorithm for noise-degraded speech. We propose a new circular average magnitude sum function (CAMSF) and a pseudo normalized correlation function (PNCF) both of which exhibit the periodicity at the pitch period of voiced speech. Exploiting the fact that CAMSF produces a peak while PNCF shows a notch, an integrated time-domain function (ITDF) is developed to enhance the pitch-harmonic-notches in presence of noise. Moreover, a frequency-frame relative smoothed noisy spectrum that acts as a harmonic spectral structure enhancer is utilized to accurately acquire a pitch-harmonic (PH) from noisy speech. We argued that employing the PH, pitch information can be effectively extracted through a variable-period impulse-train in conjunction with the proposed ITDF. It has been ascertained that the overall algorithm simulated using the Keele reference database is able to outperform some of the existing methods and well suited for a wide range of signal-to-noise ratios (SNRs) upto-10 dB.

As developments in wireless networks continue, there is an increasing expectation with regard to supporting high- quality real-time video streaming service in such networks. The recent advances in multi-description (MD) multi-path transport has made it a promising technology for content-rich wireless multimedia communications. This paper presents a rate-distortion (R-D) optimized packet scheduling algorithm (OPT- MD) for streaming MD-coded video along multiple wireless paths. Our algorithm relies on R-D hint information that is used to characterize a packet in a R-D sense. The information consists of the size of the packet in bits and the importance of the packet for reconstructing the video. Each of the video description adaptively selects certain important packets for transmission according to the quality of the transmission path by simultaneously considering bandwidth, bit error rate, and delay so that the overall end-to-end video distortion in terms of the mean square error (MSE) is minimized. Extensive simulation results demonstrate that OPT-MD can improve the quality of video streaming significantly as compared to a conventional scheduling approach that does not consider the relative importance of the video packets and the channel conditions (RANDOM-MD). The gains in performance reach up to 5 dB and 4 dB for streaming MD-coded format QCIFFORMANandTABLEvideo sequences, respectively, in the scenario of adaptation to a simulated time- varying network channel. Our efforts in this work provides an important methodology for high-quality real-time video streaming applications over wireless networks.

Statistical estimation techniques for the wavelet-based image denoising use suitable probability density functions (PDFs) as prior functions for the image coefficients. Due to the intrascale dependency of the local neighboring image wavelet coefficients, the prior functions are assumed to be stationary. In this paper, it is shown that the stationary Gram-Charlier (GC) PDF models the image coefficients better than the traditional ones, such as the stationary Gaussian and stationary generalized Gaussian PDFs. A Bayesian wavelet-based maximum a posteriori estimator is then developed by using the proposed GC prior function. Experimental results on standard images show that the proposed estimator provides a denoising performance, which is better than that of several existing denoising methods in terms of signal-to-noise ratio and visual quality.

Medical ultrasound images are widely used for diagnostic purposes. A major problem regarding these images is in their inherent corruption by speckle noise in a multiplicative fashion. The presence of speckle noise severely hampers the interpretation and analysis of medical ultrasound images. This paper presents a fast and reliable wavelet-based method for reducing the speckle in medical ultrasound images. A wavelet-based Bayesian maximum a posteriori denoiser is developed in a homomorphic framework. The wavelet coefficients of the log-transformed signal are modelled by a conditional Gaussian distribution, whereas those of the log-transformed speckle with a Maxwell distribution. The signal variances are obtained by using the local neighbors thus, making the method spatially adaptive. Simulations are performed using synthetically speckled and real ultrasound images. The results show that the proposed method can perform better than some of the existing methods in terms of the signal-to-noise ratio. Furthermore, the proposed method is fast, and preserves diagnostically important details.

In this paper, an identification technique for the minimum-phase autoregressive moving average (ARMA) systems using only the noise-corrupted observations is presented. In order to obtain a more accurate estimate of the AR parameters in the noisy environment, a repeated autocorrelation function (RACF) of the observed data is employed in the modified least-squares Yule-Walker equations. It has been found that at a very low signal-to-noise ratio (SNR), the effect of the additive noise can be significantly reduced if a twice-RACF is employed instead of the conventional ACF. Prior to the MA part identification, a noise-compensation scheme is proposed which operates on the noise-contaminated residual signal. The MA parameters are extracted from the noise-compensated power spectrum of the residual signal using the spectral factorization. ARMA systems of different orders and some natural speech signals are tested and computer simulations demonstrate a superior identification results even at a very low SNR.

A great deal of research has been conducted in the area of scheduling DSP data flow graphs (DFG) onto multiprocessor systems. This paper introduces a new scheduling and allocation algorithm for the synthesis of DSP applications. The proposed technique provides the designer with a greater flexibility to explore the design space using a hybrid arithmetic functional unit library composed of both fixed operation-specific units and reconfigurable functional units capable of executing multiple operations. A novel reconfigurable multiplier called morphable multiplier is incorporated in the proposed synthesis technique. We show that moving from a fully homogenous multiprocessor design using fixed multiple-operation units (i.e., ALUs) to a fully heterogeneous design using fixed operation-specific units (i.e., adders or multipliers) results in decreasing the area of the design, but increasing the inter-processor communication overhead. However, a hybrid multiprocessor architecture that uses a hybrid arithmetic functional unit library composed of both fixed operation-specific units and run-time reconfigurable multiple-operation units brings about a trade-off between the area and the inter-processor communication overhead

Reducing the leakage power in embedded SRAM memories is critical for low-power applications. Raising the source voltage of SRAM cells in standby mode reduces the leakage currents effectively. However, in order to preserve the state of the cell in standby mode, source voltage cannot be raised beyond a certain level. The maximum source voltage of an SRAM cell is determined by its hold stability in a particular process corner. Hence, in order to achieve the maximum leakage reduction, the source voltage of each individual cell must be raised up to its maximum safe level. However, any cell-based technique realizing this would be practically not feasible. In this paper, we propose an SRAM leakage reduction technique, referred to as adaptive sleep transistor biasing, which automatically fine-tunes the source voltage of individual memory blocks to their optimum level. Thus, maximum leakage savings can be expected while data is safely retained during standby mode. Preliminary study shows that the proposed scheme has the potential of providing substantial saving in leakage power over those by using the conventional techniques.

A two-step algorithm for the voiced/unvoiced (V/UV) decision of colored noise-corrupted speech is presented in this paper. An effective noise-whitening process is first applied to the noisy speech to combat the adverse effect of colored noise. Then, a harmonicity measure is proposed which is derived from the LP residual of the pre-whitened speech. Integrating root-mean-square energy and zero-crossing rate, another composite measure is introduced. In the first step, signal-dependent initial-thresholds (SDITs) for both the measures which are capable of highlighting distinctive attributes of voiced and unvoiced frames, are determined analyzing their statistical properties. In the second step, based on the SDITs, a bi-feature logical target function is formulated to attain a preliminary score of V/UV decision. Additional voicing criteria are developed to conquer the artifacts that may exist due to the overlapping between decision regions. Simulation results demonstrate that the proposed algorithm yields superior performance in comparison with some of the existing V/UV decision schemes in the same interfering colored noise scenario.

In this paper, an efficient VLSI architecture for a fast computation of the 2-D discrete wavelet transform (DWT) is proposed. The architecture employing a three-stage cascade in pipeline mode enhances the computing time by appropriately distributing the overall computational load among the three stages and by incorporating parallelism at various hierarchies of the architecture. The computing time is further enhanced by making use of a scheme for the equalization of the data paths in terms of the delays in the computational blocks of the architecture. A Verilog simulation of the proposed architecture is carried out to demonstrate the superior performance of the architecture.

The speckle noise complicates the human and automatic interpretation of synthetic aperture radar (SAR) images. Thus, the reduction of speckle is critical in various SAR image processing tasks. In this paper, we introduce a new spatially adaptive wavelet-based Bayesian method for despeckling the SAR images. The wavelet coefficients of the logarithmically transformed reflectance and speckle noise are modeled using the zero-location Cauchy and zero-mean Gaussian distributions, respectively. These prior distributions are then exploited to develop a Bayesian minimum mean absolute error estimator as well as a maximum a posteriori estimator. A new context-based technique with a reduced complexity is proposed for incorporating the spatial dependency of the wavelet coefficients with the Bayesian estimation processes. Experiments are carried out using typical noise-free images corrupted with simulated speckle noise as well as real SAR images, and the results show that the proposed method performs favorably in comparison to some of the existing methods in terms of the peak signal-to-noise ratio, speckle statistics and structural similarity index, and in its ability to suppress the speckle in the homogeneous regions

In order to take advantage of the byte-type data parallelism in the existing single-instruction multiple-data (SIMD) technique, this paper introduces the concept of 8-bit partial sums, obtained by a 4-bit right-shift operation on the sum of the 16 luminance values in a column of a 16 x 16 block of a video frame. Since these partial sums are of only eight bits, eight of them can be processed concurrently in a single 64-bit SIMD register. A method of employing these partial sums in order to speed up a given block motion-estimation algorithm is then proposed. The notion of the 8-bit partial sums is extended to the four-level case. It is shown that there are 15 possible methods of utilizing these multilevel 8-bit partial sums to accelerate a block motion-estimation algorithm without any loss of accuracy of the algorithm. Each of these 15 methods is used in the full-search algorithm to determine the one that provides the lowest computational complexity. This method is adopted as the chosen scheme to accelerate various block motion-estimation algorithms. Extensive simulations are carried out on eight video sequences showing that substantial speed-up can be achieved when the chosen scheme is incorporated with the various motion-estimation algorithms. The simulation results also demonstrate that the implementation on SIMD architectures can further accelerate the execution of the proposed scheme by more than 93% percent.

The paper proposes a joint probability density function to model the video wavelet coefficients of any two neighboring frames and then applies this statistical model for denoising. The parameter of the density function that measures the correlation between the wavelet coefficients of the two frames is used as an index for the motion. The joint density function is employed for spatial filtering of the noisy wavelet coefficients by developing a bivariate maximum a posteriori estimator. A recursive time averaging of the spatially filtered wavelet coefficients is adopted for further noise reduction. Simulation results on test video sequences show an improved performance both in terms of the peak signal-to-noise ratio and the perceptual quality compared to that of the other denoising algorithms

The problem of composite hypothesis testing where the probability law governing the generation of the free parameter is not explicitly known is considered. It is shown that unlike the Neyman-Pearson (NP) approach, the competitive NP (CNP) approach models incomplete prior information about the source into the detector design by setting a variable upper bound for the probability of false-alarm term. Further, the CNP and NP approaches are employed to develop the CNP and NP detectors for voice activity detection (VAD), where the prior SNR is shown to be the free parameter of the composite hypothesis. We test the CNP and NP detectors using speech samples from the SWITCHBOARD database which are suitably corrupted using different noises and various SNRs. Our simulation results show that the CNP detector outperforms its NP counterpart and is comparable to the adaptive multi-rate (AMR) VADs

It is very important to perceptually shape the watermark signal before embedding it into a host image according to the characteristics of the HVS (human vision system) since watermark invisibility is a necessary requirement for a successful watermarking application. Two popular HVS models have been proposed to deal with this problem: DCT-based Watson model and DWT-based PWM (pixel-wise masking) model, which correspond to the DCT-based and DWT-based watermarking techniques, respectively. Even though there is a common belief that the PWM model is better than the Watson model, there have been no studies that compare these two approaches. This paper is devoted to such a comparison. Our results show that the believed superiority of the PWM model relative to the Watson model is not correct and the Watson model indeed outperforms the PWM model. We argue that more accurate masking strategies in the wavelet domain are needed for DWT-based watermarking applications.

In this paper, we introduce a new spatially adaptive homomorphic Bayesian wavelet-based method for despeckling synthetic aperture radar (SAR) images. The wavelet coefficients of the logarithmically transformed reflectance image and the speckle noise image are modeled using a symmetric normal inverse Gaussian prior and an additive white Gaussian noise distribution, respectively. These models are then exploited to develop a Bayesian maximum a posteriori estimator. A method is proposed for estimating the parameters of the assumed prior. The noise-free variance of a wavelet coefficient is locally estimated, and used in a minimum mean square error estimator to obtain the corresponding noise-free coefficient. Experiments are carried out on two synthetically speckled images and a real SAR image. The results show that the proposed method has a performance that is superior to that of the other existing methods in terms of the peak signal-to-noise ratio, ability to suppress the speckle in the homogeneous regions.

This paper proposes an efficient algorithm to compute the reverse jacket transform algorithm by introducing a two-step decomposition strategy coupled with an appropriate use of the Kronecker product. Comparisons are carried out with the existing algorithms and the results show that a significant reduction in the number of data transfers and address generations as well as the structural complexity can be easily achieved using the proposed algorithm without increasing the arithmetic complexity. It is also shown that the three weights used in the existing reverse jacket transform are not required and just two are sufficient. Further, it is shown that a significant reduction in the number of multiplications can be achieved by using two rather than three weights, without losing the generality of the reverse jacket transform

In this paper, the existing one-dimensional (1-D) radix-2/4 decimation-in-frequency (DIF) fast Fourier transform (FFT) algorithm is generalized to the case of an arbitrary dimension by introducing a mixture of radix-(2 times 2 times ... times 2) and radix-(4 times 4 times ... times 4) index maps. The introduction of these index maps coupled with an appropriate use of the Kronecker product enable us to design an efficient multi-dimensional (M-D) split vector-radix DIF FFT algorithm and characterize its butterfly by simple closed-form expressions allowing easy software or hardware implementation of the algorithm for any dimension. It is shown that the proposed algorithm substantially reduces the complexity compared to the existing M-D FFT algorithms

In this paper, a new model for the ramp-cepstrum of the one-sided autocorrelation function of a noise-free autoregressive moving average (ARMA) signal is presented. The proposed blind identification technique can estimate the parameters of ARMA systems in both noise-free and noisy environments without using the input observations. It is shown that, utilizing the proposed ARMA ramp-cepstrum model in accordance with a residue-based least-squares optimization technique, both AR and MA parameters of ARMA systems can be directly obtained. The proposed method is tested on synthetic ARMA systems of different orders and also on some natural speech signals. Simulation results demonstrate the efficacy of the proposed identification scheme at low to high SNR levels

This paper presents a new algorithm for the voiced/unvoiced (V/UV) decision of noise-corrupted speech. A speech periodicity-harmonic function (SPHF) is proposed to manifest distinctive characteristics between voiced and unvoiced regions. A composite feature vector is developed by combining a periodicity measure obtained from the SPHF with some energy measures such as zero-crossing rate-weighted RMS energy, Kaiser-Teager frame energy and the normalized low-frequency energy ratio. Unlike the conventional hard threshold, a signal-dependent initial-threshold (SDIT) for each feature is determined based on its statistical properties. The SDIT is exploited to develop a logical expression that returns an objective score regarding V/UV region. Additional voicing criteria are introduced to remove the artifacts that may exist due to the overlapping between decision regions. Simulation results of the proposed multifeature classification scheme, using the Keele reference database, show superior efficacy at a low SNR relative to some of the existing V/UV decision algorithms

In the multiresolution motion estimation (MRME) techniques originally proposed by Zhang and Zafar, four MRME algorithms have been proposed. In one of these algorithms, the motion vectors in the low-pass subband are properly scaled and used as the final motion vectors for all the other subbands, and, in another, the properly scaled motion vectors in the first algorithm are used as predictions and further refined. The former algorithm requires a much lighter computational load and fewer coding bits for the motion vectors than the latter; on the other hand, the latter is able to provide a better MRME performance than the former. In this paper, we propose a new MRME technique that takes advantage of both of the above algorithms. In the proposed algorithm, the sum of absolute difference associated with each of the scaled motion vectors as in the first algorithm is calculated, and the result compared with the sum of the absolute values of the amplitudes of the wavelet coefficients within the motion block to be compensated. The outcome of the comparison decides if these scaled motion vectors are accepted as the final ones. For the coding of motion information, the motion vectors used for the prediction and their patterns of applicability to higher resolution levels, called the indices, are coded.

The performance of various wavelets, including those known to be well suited for the coding of still images, has been evaluated for the multiresolution motion estimation of video sequences. The multiresolution motion estimation scheme proposed by Zhang and Zafar, which has been widely cited in the literature, is used as the simulation scheme in this study. In our study, the prediction mean square error in the wavelet transform coefficient domain is used as the measure for prediction performance. In order to show the overall rate distortion performance, the number of bits needed to encode the motion vectors is also calculated. Simulation results show that the 7/9 biorthogonal wavelet, one of the best wavelets for the coding of still images, is the best wavelet for the task of multiresolution motion estimation among the wavelets evaluated in this study.

A radix-2/16 decimation-in-frequency (DIF) fast Fourier transform (FFT) algorithm and its higher radix version, namely radix-4/16 DIF FFT algorithm, are proposed by suitably mixing the radix-2, radix-4 and radix-16 index maps, and combing some of the twiddle factors. It is shown that the proposed algorithms and the existing radix-2/4 and radix-2/8 FFT algorithms require exactly the same number of arithmetic operations (multiplications+additions). Moreover, by using techniques similar to these, it can be shown that all the possible split-radix FFT algorithms of the type radix-2/sup r//2/sup rs/ for computing a 2/sup m/-point DFT require exactly the same number of arithmetic operations.

A new approach for the identification of minimum-phase autoregressive moving average (ARMA) systems in the presence of heavy noise is presented in this paper. A damped sinusoidal (DS) model for the autocorrelation function of a noise-free ARMA signal is proposed to estimate the AR parameters, which overcomes the failure of conventional correlation based techniques in estimating the AR parameters of an ARMA system at a very low signal-to-noise ratio (SNR). The MA parameters of the ARMA system are then estimated by using Durbin's method along with an optimum order selection criterion. Both white noise and periodic impulse train excitations are considered for the application of the proposed method to system identification as well as to speech processing. Computer simulations are carried out based on both synthetic ARMA systems and natural speech signals, showing superior identification results even at an SNR of -5 dB for which most of the existing methods would fail.

In this paper, we present a joint time/frequency domain approach for pitch estimation of speech at a very low SNR. The kernel of this approach lies in introducing a new function for detecting the time-domain cue by modifying the circular average magnitude difference function (CAMDF). By using the new function in conjunction with the half-wave rectified version of the autocorrelation function, the pitch-peak can be emphasized and the non-pitch peaks suppressed. To guarantee a robust pitch detection in noisy speech, a priori frequency-domain estimate of the dominant pitch-harmonic is extracted as an additional cue and is utilized to optimally match the pitch-peak in time-domain. The proposed approach is simulated using the Keele reference database. It is shown that the proposed method using joint time and frequency domain cues is able to give a superior accuracy relative to some of the existing methods even at a very low SNR of -10 dB.

We provide a framework for developing a low-complexity adaptive filtering algorithm by incorporating the concept of partial-updating into the technique of finding the gradient vector in the hyperplane based on the L/sub /spl infin//-norm criterion. The resulting algorithm is referred to as the partial-update normalized sign LMS (PU-NSLMS) algorithm. A specific case of the PU-NSLMS algorithm, called the M-Max PU-NSLMS algorithm, based on the concept of having a minimum Euclidean length of the coefficient-update vector, is considered. It is shown that this algorithm is computationally less complex compared to the partial-update normalized least-mean squares (PU-NLMS) algorithm. Results concerning the mean-square analysis of the M-Max PU-NSLMS algorithm are given. The performance of this algorithm is compared with that of the PU-NLMS algorithm in the case of network echo cancellation. It is shown that the convergence rate of the proposed algorithm is comparable to that of the PU-NLMS algorithm, but with a reduced complexity, making it a good choice for applications requiring a long filter tap, especially for real-time implementations.

In this paper, a new architecture of the redundant residue number system (RRNS) based quasi-chaotic coding is proposed for the secure telecommunication systems and networks. In the proposed architecture, a number of modulo operations required by the existing designs are replaced by binary coding operations to simplify the design. Also, a moduli selection method and a residue-to-binary converter with error-correction capability are proposed to further improve the efficiency of the design specifically for FPGA implementation. The proposed architecture is implemented and tested using Matlab and Xilinx FPGA hardware. The results show that compared to the existing design, the proposed design requires only 80% of the hardware resources while maintaining the same speed. The power consumption is also reduced by 25%.

For original paper see Pei and Chen (IEEE Signal Process. Lett., vol.11, no.5, p.459-62, 2004). The present authors report that the same algorithm has been previously proposed in Bouguezel et al. (IEEE Int. Symp. Circuits Syst., vol.3, p.698-701, 2003).

A special class of even-symmetric periodic signals is introduced. The most distinctive feature of these signals is that their real-valued Fourier coefficients can be calculated by forming a weighted average of the signal values using integer-valued coefficients. The signals arise from number-theoretic concepts concerning a class of functions called even arithmetical functions. The integer-valued weighting coefficients, being sums of complex roots of unity, are the Ramanujan sums and may be computed recursively or through closed-form arithmetical relations. The recursive method of computation is based on the cyclotomic polynomials and is described in detail. If the signal values are integers, the computation of the discrete Fourier transform (DFT) coefficients of this class of signals can be performed in an exact quantization-error-free manner by performing arithmetical operations on integers. The theoretical development is supplemented by concrete examples.

In this paper, by using the properties of the higher order statistics (HOS) of speech and noise signals, we develop an improved voice activity detection (VAD) scheme. The proposed scheme employs the logarithm of the kurtosis of the LPC residual of a speech signal and is shown to be more effective and efficient in detecting active speech in medium to low signal-to-noise ratio (SNR) conditions without being unduly affected by the variations in the signal energy. To overcome the inability of the HOS in detecting unvoiced speech, another metric (the low band to full band energy ratio) is introduced. Depending on the estimated mean SNR, the proposed scheme works adaptively in two modes: a simple mode using only the SNR, and an enhanced mode using the HOS, the low band to full band energy ratio and the SNR. This scheme is capable of avoiding unnecessary computations, while maintaining the same performance as that working only in the enhanced mode. Simulations results are presented to demonstrate the effectiveness of the proposed voice activity detection scheme.

In this paper, we present a method of kernel optimization by maximizing a measure of class separability in the empirical feature space, an Euclidean space in which the training data are embedded in such a way that the geometrical structure of the data in the feature space is preserved. Employing a data-dependent kernel, we derive an effective kernel optimization algorithm that maximizes the class separability of the data in the empirical feature space. It is shown that there exists a close relationship between the class separability measure introduced here and the alignment measure defined recently by Cristianini. Extensive simulations are carried out which show that the optimized kernel is more adaptive to the input data, and leads to a substantial, sometimes significant, improvement in the performance of various data classification algorithms.

In this paper, we propose a flexible architecture that performs the computation of the discrete wavelet transform, requiring a small memory space and is capable of operating at high sampling rate. The architecture employs two filtering blocks to compute the transform and one buffer to store the intermediate results. Each filtering block has two processing units that operate independently in parallel using a two-phase scheduling. An efficient scheme for the synchronization of the data flow among the three blocks is provided in order to minimize the buffer size and increase the speed of operation. Verilog and HSPICE simulation results are presented to show that the proposed architecture is more efficient for the computation of a fully decomposed discrete wavelet transform with high-tap filters than some other existing architectures in terms of their areas and speed of operations.

In this paper, we propose a low cost motion-compensated interpolation technique to improve the video quality for the low bit-rate video encoded in conjunction with frame dropping. The proposed approach exploits the block-based motion vector field available to the decoder to avoid the complex motion estimation at the receiver. An iterative refinement technique derived using the finite element method is employed to efficiently conceal the interpolation errors caused by unfilled and overlapped pixels in the predicted frames. Consequently, no pixel classification is needed in the proposed technique, thus substantially reducing the computational complexity. Simulation results show that this technique results in reconstructed frames with good visual quality.

A discrete-time fractional-order differentiator is modeled as a finite-impulse response (FIR) system. The system yields fractional-order derivatives of Riemann-Liouville type for a uniformly sampled polynomial signal. The computation of the output signal is based on the additive combination of the weighted outputs of N cascaded first-order digital differentiators. For differentiators of fractional order with a terminal value equal to zero, the weights are time-varying. The weights are obtained in a closed form involving the Stirling numbers of the first kind. The system tends to a time-invariant integer-order differentiator when the order of the derivative tends to an integer value. It yields exact fractional- or integer-order derivatives of a sampled polynomial signal of a certain order.

Codebook initialization usually has a significant effect on the performance of vector quantization algorithms. This letter presents a new scheme of codebook initialization in which the competitive learning and code vector splitting are incorporated together to produce a good initial codebook. Based mainly on the geometrical measurements of the learning tracks of the code vectors, the competitive splitting mechanism shows an ability to appropriately allocate code vectors according to the spatial distribution of the input data and, therefore, tends to give a better initial codebook. Comparisons with other initialization techniques demonstrate the effectiveness of the new scheme.

It is shown that, by choosing appropriate values for the parameter a of the generating kernel in constructing P.J. Burt and E.H. Adelson's pyramid for motion estimation (see IEEE Trans. Commun., vol.31, p.337-45, 1983), one can eliminate the operation of the floating point multiplications needed in such a construction and reduce the computational load to the same order as in the case of the mean pyramid. When a is chosen to be 3/8, it is demonstrated, through simulation studies, that the corresponding Burt and Adelson's pyramid does not degrade the performance of motion estimation, as compared to that using kernels giving the best performance, and this pyramid also provides perceptually better motion-compensated images than those provided by the mean pyramid.

This paper presents a new self-creating model of a neural network in which a branching mechanism is incorporated with competitive learning. Unlike other self-creating models, the proposed scheme, called branching competitive learning (BCL), adopts a special node-splitting criterion, which is based mainly on the geometrical measurements of the movement of the synaptic vectors in the weight space. Compared with other self-creating and nonself-creating competitive learning models, the BCL network is more efficient to capture the spatial distribution of the input data and, therefore, tends to give better clustering or quantization results. We demonstrate the ability of the BCL model to appropriately estimate the cluster number in a data distribution, show its adaptability to nonstationary data inputs and, moreover, present a scheme leading to a multiresolution data clustering. Extensive experiments on vector quantization of image compression are given to illustrate the effectiveness of the BCL algorithm.

Fast block motion estimation algorithms are needed for real-time implementations of video coding standards due to the high computational complexity of the full-search algorithm for block motion estimation. In this paper, an algorithm using 8-bit partial sums of 16 luminance values for a fast block motion estimation is proposed. The technique of using the partial sums is employed to reduce the computational complexity of not only the full-search algorithm but also some of the fast block motion estimation algorithms while maintaining their accuracy. Furthermore, it is shown that the byte-type data-parallelism on an SIMD architecture can be utilized to access and process these partial sums concurrently to accelerate the process of motion estimation. Simulation results are presented to demonstrate that the use of the partial sums can accelerate the execution of the full-search, three-step search, and four-step search algorithms on an SIMD architecture significantly.

In this paper, a pyramidal motion estimation technique that makes use of the motion correlation within a pyramidal level is proposed. Instead of scaling the motion vectors from the adjacent lower pyramidal level as motion predictions as is done in the conventional technique, in the proposed technique, the motion vectors from the neighboring motion blocks are taken into consideration as possible candidates. Each of these candidate motion vectors is used for prediction and refined. The motion vector that has the least matching distortion is chosen as the final motion vector. Compared to the conventional pyramidal motion estimation technique, the proposed method effectively overcomes the problem of propagation of false motion vectors. Simulation studies show that a substantial performance improvement is achieved, both in terms of the prediction mean square error and the number of coding bits for the motion vectors.

In this paper, a non-causal median filtering method is proposed to predict the motion vectors across the wavelet subbands of a video frame for multiresolution motion estimation. This median filtering method effectively overcomes the problem of propagation of false motion vectors that exists in the conventional multiresolution motion estimation schemes. A significant feature of the proposed technique is that it imposes no demand for additional bandwidth. Simulation studies show that this median filtering-based multiresolution motion; estimation technique effectively improves the motion prediction performance. It is further shown that this performance improvement is achieved with little increase in the computational complexity.

This paper introduces a novel technique to obtain a schedule for a cyclic data-flow graph (DFG) onto a multiprocessor system. The optimality criteria considered in this scheduling technique are the maximum throughput, minimum input–output (I/O) delay, and minimum hardware resources. In this technique, an all-pair longest path algorithm is used to evaluate the relative firing times of the nodes of the given DFG. The proposed technique for finding these times is quite simple to implement and it has lower time complexity than all the previously proposed techniques. The technique is tested on various benchmark problems to demonstrate its optimal performance. All the optimality criteria are achieved on all the tested benchmarks. However, finding a minimum hardware resource schedule is an NP complete problem, and thus cannot be theoretically ensured. A formal proof of achieving both the throughput and the I/O delay optimality simultaneously is given, and an efficient technique to ensure this is also presented. This technique is quite simple and can be used to ensure delay optimality in any scheduling technique.

We investigate three new methods to predict motion vectors (MVs) across subbands for multi-resolution motion estimation (MRME): linear prediction, median filtering (MF), and multi-candidate techniques. Compared to the conventional MRME techniques, the proposed linear-prediction-based and the MF-based techniques effectively overcome the problem of propagation of false MVs. A significant feature of these two techniques is that they impose no demand for additional bandwidth, and simulation studies show that they not only improve prediction performance, but also reduce the number of bits needed to encode the motion information. It is further shown that the improvement in the performance thus achieved involves little increase in computational complexity.

A median filtering-based hierarchical motion vector estimation scheme making use of a pyramidal data structure is proposed. Compared to the conventional hierarchical motion vector estimation schemes, the proposed scheme overcomes the problem of propagation of false motion vectors across resolutions. Simulation studies show that the proposed scheme not only improves the prediction accuracy with respect to the prediction mean square error, but also results in a smoother motion field, which can be encoded with less number of coding bits. It is shown that an improvement in the rate distortion performance is achieved with little increase in the computational complexity. It is also shown that Burt and Adelson's (1983) pyramidal data structure provides the best performance among a number of the generating kernels considered in our study.

A well known result of rate-distortion theory states that, under broad conditions, the quantization error has a Gaussian distribution. It is also known that a Gaussian memoryless source is successively refinable. These results indicate that the use of code books designed for a generic Gaussian source for different stages of a residual vector quantizer does not result in loss of performance. In this work, we present a residual vector quantizer using an optimal (LBG) vector quantizer in the first stage and a Gaussian codebook in the other stages. The closeness of the distribution of the error signals to the Gaussian distribution is examined and it is shown that while the rate-distortion theoretic results are true only when the rate of the first stage is very high, in practice, even at moderate rates, the loss in the optimality is quite small.

Based on the mechanisms underlying the performance of fractal and DWT, one method using fractal-based selfquantization coding approach to code different subband coefficients of DWT is presented. Within this method finer coefficients are fractal encoded according to the successive coarser ones. Self-similarities inherent between parent and their children at the same spatial location of the adjacent scales of similar orientation are exploited to predict variation of information across wavelet scales. On the other hand, with respect to the HVS model, we assign different error thresholds to different scales and different shape of range blocks to different orientations of the same scale, by which the perceptually lossless high compression ratio can be achieved and the matching processing can be quickened dramatically.

A high-speed parallel residue-to-binary converter is proposed for the moduli set S/sup k/={2/sup m/-1, 2/sup 2(0)m/+1, 2/sup 2(l)m/+1, L, 2.

The conventional variable-size block motion compensation technique, even though superior to the fixed-size block motion-compensation technique, cannot fully utilize the motion information of a frame for its partitioning. This paper presents a quadtree structured region-wise motion-compensation technique that utilizes more effectively the motion content of a frame in terms of the shape, size, and location of the partitioned regions. The proposed technique is based on a new coding scheme of the quadtree structure, where a two-bit code is used. The partitioning of a given frame is carried out through the processes of merging and combining. The merging process partitions the frame into a number of regions by carrying out all possible 4-to-1, 3-to-1, and 2-to-1 merges along the quadtree hierarchy according to some specified criteria, while the combining process combines some of the partitioned regions that have the same motion vector into a single region. The motion vectors of the resulting partitioned regions are coded by a new coding scheme that uses the temporal correlation of the motion fields. Test results of applying the proposed method on a number of MPEG video sequences are included. These results indicate that the proposed method can provide a significantly improved rate-distortion performance.

This paper addresses the problem of designing optimal stack filters by employing an Lp norm of the error between the desired signal and the estimated one. It is shown that the Lp norm can be expressed as a linear function of the decision errors at the binary levels of the filter. Thus, an Lp-optimal stack filter can be determined as the solution of a linear program. The conventional design of using the mean absolute error (MAE), therefore, becomes a special ease of the general Lp norm-based design developed here. Other special cases of the proposed approach, of particular interest in signal processing, are the problems of optimal mean square error (p=2) and minimax (p-->infinity) stack filtering. Since an Linfinity optimization is a combinatorial problem, with its complexity increasing faster than exponentially with the filter size, the proposed Lp norm approach to stack filter design offers an additional benefit of a sound mathematical framework to obtain a practical engineering approximation to the solution of the minimax optimization problem. The conventional MAE design of an important subclass of stack filters, the weighted order statistic filters, is also extended to the Lp norm-based design. By considering a typical application of restoring images corrupted with impulsive noise, several design examples are presented, to illustrate the performance of the Lp-optimal stack filters with different values of p. Simulation results show that the Lp-optimal stack filters with p=or>2 provide a better performance in terms of their capability in removing impulsive noise, compared to that achieved by using the conventional minimum MAE stack filters.

Recently, it has been shown that radix-2 DFT algorithms can be designed based on vector representation of data providing several advantages. In this paper, an index mapping approach is used to derive these algorithms. This approach makes the derivation simpler and provides a better insight into the functioning of the algorithms. The computational complexity and the performance of the software implementation of the algorithms are also presented.

The discrete Walsh and Hadamard transforms are often used in image processing tasks such as image coding, pattern recognition, and sequency filtering. A new discrete Walsh transform (DWT) algorithm is derived in which a modified form of the DWT relation is decomposed into smaller-sized transforms using vectorized quantities. A new sequency-ordered discrete Hadamard transform (DHAT) algorithm is also presented. The proposed approach results in more regular algorithms requiring no independent data swapping and fewer array-index updating and bit-reversal operations. An analysis of the computational complexity and the execution time performance are provided. The results are compared with those of the existing algorithms.

This paper addresses the problem of designing weighted order statistic (WOS) filters by employing an objective function given as the L/sub p/ norm of the error between the desired signal and the estimated one. The conventional design of WOS filters uses a mean absolute error (MAE) objective function, and as such, it is a special case of the general, L/sub p/ norm based design, developed here. It is shown that in stack filtering, the L/sub p/ norm can be expressed as a linear combination of the decision errors incurred by the Boolean operators at each level of the stack filter architecture. Based on this formulation of the L/sub p/ norm, both nonadaptive and adaptive algorithms for the design of L/sub p/ WOS filters are developed. A design example is considered, to illustrate the performance of the designed L/sub p/ WOS filters with different values of p. The simulation results show that the L/sub p/ WOS filters with p/spl ges/2 are capable of removing more impulsive noise compared with the conventional MAE WOS filters.

The variable size block motion compensation (VSBMC) technique is known to be more effective than the fixed-size block motion compensation technique (FSBMC) for video coding. However, the existing VSBMC techniques do not fully utilize the motion information of a frame for its partitioning. This paper presents a new adaptive partitioning scheme that utilizes more effectively the motion content of a frame in terms of the shape and size of the blocks. A new tree structure is proposed that allows not only the conventional four-to-one merge, but also three-to-one and two-to-one merges. This new VSBMC method results in a significantly improved rate-distortion performance. Test results of applying the proposed method on some MPEG video sequences are included.

The conventional fixed-size block motion compensation technique has two major drawbacks. One is the blocking effect that is visually annoying along with a residual image that is difficult to compress. The other is an ineffective and inefficient representation of the motion information because of the non-adaptive frame partitioning. This paper presents a new method that can adaptively divide a frame according to its motion contents and also effectively reduce the blocking effect by applying a windowing technique. Simulation results show that the proposed method can significantly improve the visual quality of the prediction image and increase the coding efficiency.

In this paper, we investigate the application of multidimensional IIR digital filters for video signal processing. The problems of sampling structure conversion and deinterlacing are addressed. A multistage filter structure is then proposed to perform the general conversions among different types of sampling structures. Some simple IIR digital filters are proposed and simulated particularly for the above applications. Some simulation results are also given.

This paper is concerned with the compile-time (static) scheduling of data flow graphs (DFGs) onto multiprocessor systems. It mainly concentrates on producing a rate-optimal time schedule that achieves the minimum iteration period, known as the iteration period bound. A combinatorial theory is developed to produce a rate-optimal time schedule for a fully specified DFG. The DFG is first converted to a critical graph by making all its circuits critical. Next, it is transformed into an acyclic graph through a sequence of circuit contractions. An algorithm is then proposed which achieves the time scheduling of the given DFG by first scheduling the acyclic graph, followed by a scheduling of the critical circuits in an order which is reverse to that of their contraction.

A Hebbian-type learning algorithm for the total least-squares parameter estimation is presented. An asymptotic analysis is carried out to show that the algorithm allows the weight vector of a linear neuron unit to converge to the eigenvector associated with the smallest eigenvalue of the correlation matrix of the input signal. When the algorithm is applied to solve parameter estimation problems, the converged weights directly yield the total least-squares solution. It is shown that the implementations of the proposed algorithm have the simplicity of those of the LMS algorithm, but its noise rejection capability is much superior to those of the least-squares-based algorithms. The applicability and performance of the algorithm are demonstrated through computer simulations of adaptive FIR and IIR parameter estimation problems.>

Least-squares (LS) estimation with a standard feedback neural network (SFBNN) which is based on an electrical model is investigated. In the energy function of a SFBNN, a non-quadratic term is included which is often neglected while solving an optimization problem. It is shown that the non-quadratic term affects the solution of a continuous optimization problem. Properties of the non-quadratic term and the relation between the estimation error and several parameters of the SFBNN are discussed. A technique, called extended space iterative search (ESIS), is introduced to reduce the estimation error. Simulation results are presented to confirm the analysis result and the effectiveness of the proposed technique.>

Problems in which the arguments of objective functions are real numbers are considered. Based on the concept of the Hopfield network, a neural network that solves the least-square estimation problem is derived. With this network, the objective function can converge to any inner point of a hypercube, giving a real-valued solution with very great speed. Because of the convex nature of the chosen energy function, the problem of convergence to a local minimum does not arise. Also introduced is a space iterative search technique for finding the optimum solution that can exist at any point within the space. Finally, simulation results are given for solving problems of linear systems and parameter estimations

In this paper software and hardware design of a tightly-coupled multimicroprocessor system with distributed common memory and private memory modules, to compute the values of auto- and cross-correlation functions, to recover signals buried in noise and to compute cross- and auto-power spectral density at 128 points simultaneously on the time delay axis or frequency axis, are described. This multimicroprocessor system has two 16-bit micro-processor boards, a single common bus, a centralized common-memory (CMO) and necessary arbitration logic circuits. To each microprocessor board, a distributed common memory, a private memory, I/O and common-memory-access control circuits are added. The addressing scheme is designed such that any common-memory cell has the same address for all the microprocessors. These software and hardware mechanisms suit the present signal processing application well. The computed results are displayed on an oscilloscope or X-Y recorder.

A method is presented for the design of a class of two-dimensional (2-D) stable digital filters satisfying prescribed magnitude of constant group delay specifications. The design method generates a 2-D digital transfer function which is a product of two transfer functions, H1(z1,z2) and H2(z1,z2), corresponding to a recursive filter and a nonrecursive filter, respectively. Component H1(z1,z2) ensures a wave-digital realization, that is, the design method guarantees the generation of a corresponding analog function H1A(s1,s2) which is realizable as the transfer function of a doubly-terminated two-variable lossless network. Thus the design technique ensures that not only a given frequency response is achieved but also the generated transfer function is realizable as a cascade of a wave-digital filter and a nonrecursive digital filter. The class of filters considered here is one in which the doubly-terminated analog network used to realize the wave digital filter is a cascade s1-and s2-variable lossless two-ports with all of their transmission zeros at infinity.

In this paper, necessary and sufficient conditions are given, under which a two-variable rational function can be realized as the transfer, function of a doubly-terminated cascade of s1and s2-variable lossless two-ports, each two-port with all of its transmission zeros at the origin or infinity. Using the above conditions, transfer function of a 2-variable analog reference filter is generated. Parameters of this transfer function are used as parameters of optimization to minimize the least mean square error between the amplitude response of the desired and designed filters. It is shown that the method can also be modified to incorporate the phase and magnitude response of the 2-D filters. To illustrate the method, examples are given.

Necessary and sufficient conditions are obtained for the realization of a class of two-variable analog transfer functions. The transfer function is realized as a resistively-terminated two-port consisting of a cascade of P1and P2-variable lossless two-ports each having all of its transmission zeros at Pi=0 or pi=∞ (i=1,2), It is shown how these analog realizations may be used to realize two-dimensional wave digital filters.