Deep neural networks (DNNs) often suffer performance drops when deployed in environments that differ from the conditions they were trained on—a challenge known as domain shift. Domain Generalization (DG) seeks to address this issue by training models across multiple source domains to learn representations resilient to distributional variations and capable of generalizing to unseen target domains. Although many existing DG methods reduce domain-specific variations by aligning source distributions in a shared feature space, recent trends show that explicitly disentangling latent features into domain-invariant and domain-specific components can improve generalization.This work proposes a novel framework, called Feature Orthogonalization via Contrastive Upper-bound Separation (FOCUS) to address the domain shift problem. FOCUS encourages the separation of task-specific and domain-specific representations by leveraging domain-specific features. Furthermore, FOCUS enhances the separation by minimizing the mutual dependence between domain-agnostic and domain-specific features. We employ the Contrastive Log-ratio Upper Bound (CLUB) estimator to minimize the mutual dependence, which provides an upper bound on the mutual information. During training, FOCUS effectively enforces disentanglement in the latent space by reducing this bound and improving generalization to unseen domains.Extensive experiments on widely used DG benchmarks, including PACS, Office-Home, VLCS, and TerraIncognita, demonstrate that FOCUS consistently outperforms several state-of-the-art approaches with an average accuracy improvement of 1.03%, 0.73%, 1.14% and 2.03%, respectively. These findings highlight the importance of mutual information-based disentanglement in enhancing the generalization capabilities of deep networks to address domain shifts in real-world problems.

Large Language Models (LLMs) have gained attention in research and industry, aiming to streamline processes and enhance text analysis performance. Thematic Analysis (TA), a prevalent qualitative method for analyzing interview content, often requires at least two human experts to review and analyze data. This study demonstrates the feasibility of LLM-Assisted Thematic Analysis (LATA) using GPT-4 and Gemini. Specifically, we conducted semi-structured interviews with 14 researchers to gather insights on their experiences generating and analyzing Online Social Network (OSN) communications datasets. Following Braun and Clarke's six-phase TA framework with an inductive approach, we initially analyzed our interview transcripts with human experts. Subsequently, we iteratively designed prompts to guide LLMs through a similar process. We compare and discuss the manually analyzed outcomes with responses generated by LLMs and achieve a cosine similarity score up to 0.76, demonstrating a promising prospect for LATA. Additionally, the study delves into researchers' experiences navigating the complexities of collecting and analyzing OSN data, offering recommendations for future research and application designers.

The development of non-invasive markers for assessing the risk of sudden cardiac death has gained significant attention, particularly T-wave alternans (TWA), which can be recorded from surface electrocardiogram (ECG) signals. However, the clinical application of TWA remains insufficiently standardized, complicating its detection in real-world ambulatory environments due to variable conditions that often affect ECG recordings, including dynamic changes, noise, and artifacts. This study presents a Deep Learning (DL) approach designed to detect TWA directly from ECG signals, using transfer learning with robust architectures such as VGG, ResNet, and Inception. Our method simplifies the detection pipeline by eliminating the need for prior signal processing steps such as R-peak identification, T-wave segmentation, or feature engineering. Our models are trained on a custom long-term dataset of real patients, capturing TWA episodes ranging from non-visible micro-alternans to higher amplitude TWA of 20 to 100 $\mu$V, and incorporating a robust methodology that emphasizes patient separation during training and testing to enhance generalizability. The results demonstrate that our model achieves an F1-score of 0.83 during ambulatory analysis, outperforming traditional Machine Learning approaches. By eliminating the need for extensive preprocessing, our approach not only enhances the adaptability of TWA detection but also brings the model closer to practical applicability in clinical settings, leading to more efficient and effective risk stratification for sudden cardiac death.

Detecting malicious software applications is crucial in protecting user data and privacy. The increasing popularity of operating systems (OS) such as Android, iOS, and Windows has led to an expansion in the number of malicious software applications targeting these platforms. Several studies have explored various techniques to detect malware threats, including signature-based detection, behavioral analysis, dynamic analysis, and static analysis. However, these conventional methods may fail to detect new variants of malware with unique characteristics. This research paper introduces a unified approach that combines entropy bytecode-to-image visualization with ensemble learning to enhance the identification of malicious patterns. To validate the efficacy of the entropy encoder, the AndroZoo-2011 application data set was used as a training set, comprising 184,474 samples, including 49,150 and 13,324 malicious and benign applications, respectively. The applied method demonstrates improved performance with an accuracy of 95.77%, a precision of 90.60%, and an F1 score of 92.21% compared to other existing encoders, making our approach more applicable to real-world scenarios.

Engagement detection is essential in many areas such as driver attention tracking, employee engagement monitoring, and student engagement evaluation. In this paper, we propose a novel approach using attention based hybrid deep models for the 8th Emotion Recognition in the Wild (EmotiW 2020) Grand Challenge in the category of engagement prediction in the wild EMOTIW2020. The task aims to predict the engagement intensity of subjects in videos, and the subjects are students watching educational videos from Massive Open Online Courses (MOOCs). To complete the task, we propose a hybrid deep model based on multi-rate and multi-instance attention. The novelty of the proposed model can be summarized in three aspects: (a) an attention based Gated Recurrent Unit (GRU) deep network, (b) heuristic multi-rate processing on video based data, and (c) a rigorous and accurate ensemble model. Experimental results on the validation set and test set show that our method makes promising improvements, achieving a competitively low MSE of 0.0541 on the test set, improving on the baseline results by 64%. The proposed model won the first place in the engagement prediction in the wild challenge.

A graph neural network (GNN) for image understanding based on multiple cues is proposed in this paper. Compared to traditional feature and decision fusion approaches that neglect the fact that features can interact and exchange information, the proposed GNN is able to pass information among features extracted from different models. Two image understanding tasks, namely group-level emotion recognition (GER) and event recognition, which are highly semantic and require the interaction of several deep models to synthesize multiple cues, were selected to validate the performance of the proposed method. It is shown through experiments that the proposed method achieves state-of-the-art performance on the selected image understanding tasks. In addition, a new group-level emotion recognition database is introduced and shared in this paper.

Social relations reveal the interpersonal association of human beings. Developing techniques to automatically recognize social relations from visual data has great potential for improving human-computer interaction. In this paper, a hybrid deep network is proposed to predict the social relations between two human beings in an image. Unlike existing methods that typically learn deep learning models from scratch, a VGG-FACE model previously trained for face recognition is fine-tuned on a social relation database and used as branches of a siamese-like network. Moreover, a deep network is proposed to extract scene features that contain high-level information related to social relations from whole images and its predictions are fused with the predictions of the siamese network to generate the final result. Experiments show that the proposed approach saves the effort of pre-training and preparing auxiliary datasets, i.e. facial attribute datasets, and outperforms state-of-the-art methods.

Conformal prediction uses the degree of strangeness (nonconformity) of new data instances to determine the confidence values of new predictions. We propose an inductive conformal predictor for sparse coding classifiers, referred to as ICP-SCC. Our contribution is twofold: first, we present two nonconformity measures that produce reliable confidence values; second, we propose a batch mode active learning algorithm within the conformal prediction framework to improve classification performance by selecting training instances based on two criteria, informativeness and diversity. Experiments conducted on face and object recognition databases demonstrate that ICP-SCC improves the classification accuracy of state-of-the-art dictionary learning algorithms while producing reliable confidence values.

Group cohesiveness is a compelling and often studied composition in group dynamics and group performance. The enormous number of web images of groups of people can be used to develop an effective method to detect group cohesiveness. This paper introduces an automatic group cohesiveness prediction method for the 7th Emotion Recognition in the Wild (EmotiW 2019) Grand Challenge in the category of Group-based Cohesion Prediction. The task is to predict the cohesive level for a group of people in images. To tackle this problem, a hybrid network including regression models which are separately trained on face features, skeleton features, and scene features is proposed. Predicted regression values, corresponding to each feature, are fused for the final cohesive intensity. Experimental results demonstrate that the proposed hybrid network is effective and makes promising improvements. A mean squared error (MSE) of 0.444 is achieved on the testing sets which outperforms the baseline MSE of 0.5.

In order to reduce overfitting for the image recognition application, this paper proposes a novel regularization learning algorithm for deep learning. Above all, we propose a novel probabilistic representation for explaining the architecture of Deep Neural Networks (DNNs), which demonstrates that the hidden layers close to the input formulate prior distributions, thus DNNs have an explicit regularization, namely the prior distributions. Furthermore, we show that the backpropagation learning algorithm is the reason for overfitting because it cannot guarantee precisely learning the prior distribution. Based on the proposed theoretical explanation for deep learning, we propose a novel regularization learning algorithm for DNNs. In contrast to most existing regularization methods reducing overfitting by decreasing the training complexity of DNNs, the proposed method reduces overfitting through training the corresponding prior distribution in a more efficient way, thereby deriving a more powerful regularization. Simulations demonstrate the proposed probabilistic representation on a synthetic dataset and validate the proposed regularization on the CIFAR-10 dataset.

Smile detection from unconstrained facial images is a specialized and challenging problem. As one of the most informative expressions, smiles convey basic underlying emotions, such as happiness and satisfaction, and leads to multiple applications, such as human behavior analysis and interactive controlling. Compared to the size of databases for face recognition, far less labeled data is available for training smile detection systems. This paper proposes an efficient transfer learning-based smile detection approach to leverage the large amount of labeled data from face recognition datasets and to alleviate overfitting on smile detection. A well-trained deep face recognition model is explored and fine-tuned for smile detection in the wild, unlike previous works which use either hand-engineered features or train deep convolutional networks from scratch. Three different models are built as a result of fine-tuning the face recognition model with different inputs, including aligned, unaligned and grayscale images generated from the GENKI-4K dataset. Experiments show that the proposed approach achieves improved state-of-the-art performance. Robustness of the model to noise and blur artifacts is also evaluated in this paper.

This paper presents a hybrid deep learning network submitted to the 6th Emotion Recognition in the Wild (EmotiW 2018) Grand Challenge [9], in the category of group-level emotion recognition. Advanced deep learning models trained individually on faces, scenes, skeletons and salient regions using visual attention mechanisms are fused to classify the emotion of a group of people in an image as positive, neutral or negative. Experimental results show that the proposed hybrid network achieves 78.98% and 68.08% classification accuracy on the validation and testing sets, respectively. These results outperform the baseline of 64% and 61%, and achieved the first place in the challenge.

This paper presents the work submitted to the Group-level Emotion Recognition sub-challenge, which is part of the 5th Emotion Recognition in the Wild (EmotiW 2017) Challenge. The task of this sub-challenge is to classify the emotion of a group of people in each image as positive, neutral or negative. To address this task, a hybrid network that incorporates global scene features, skeleton features of the group, and local facial features is developed. Specifically, deep convolutional neural networks (CNNs) are first trained on the faces of the group, the whole images and the skeletons of the group, and then fused to perform the group-level emotion prediction. Experimental results show that the proposed network achieves 80.05% and 80.61% on the validation and testing sets, respectively, outperforming the baseline of 52.97% and 53.62%.

A label consistent recursive least squares dictionary learning algorithm, LC-RLSDLA, is proposed to learn discriminative dictionaries for image classification based on sparse coding. The class label information and a label consistency term are used in the cost function to enforce discriminability among the sparse codes. Two operation modes are derived for the LC-RLSDLA: the supervised learning mode, in which the algorithm employs a training set to learn the dictionary and linear classifier simultaneously, and the online semi-supervised learning mode, in which the algorithm continuously learns as it classifies new vectors. Experiments performed on two face recognition databases demonstrate that the proposed method outperforms state-of-the-art sparse coding algorithms.

Recent results in telecardiology show that compressed sensing (CS) is a promising tool to lower energy consumption in wireless body area networks for electrocardiogram (ECG) monitoring. However, the performance of current CS-based algorithms, in terms of compression rate and reconstruction quality of the ECG, still falls short of the performance attained by state-of-the-art wavelet-based algorithms. In this paper, we propose to exploit the structure of the wavelet representation of the ECG signal to boost the performance of CS-based methods for compression and reconstruction of ECG signals. More precisely, we incorporate prior information about the wavelet dependencies across scales into the reconstruction algorithms and exploit the high fraction of common support of the wavelet coefficients of consecutive ECG segments. Experimental results utilizing the MIT-BIH Arrhythmia Database show that significant performance gains, in terms of compression rate and reconstruction quality, can be obtained by the proposed algorithms compared to current CS-based methods.

Image-based classification of histology sections plays an important role in predicting clinical outcomes. However this task is very challenging due to the presence of large technical variations (e.g., fixation, staining) and biological heterogeneities (e.g., cell type, cell state). In the field of biomedical imaging, for the purposes of visualization and/or quantification, different stains are typically used for different targets of interest (e.g., cellular/subcellular events), which generates multi-spectrum data (images) through various types of microscopes and, as a result, provides the possibility of learning biological-component-specific features by exploiting multispectral information. We propose a multispectral feature learning model that automatically learns a set of convolution filter banks from separate spectra to efficiently discover the intrinsic tissue morphometric signatures, based on convolutional sparse coding (CSC). The learned feature representations are then aggregated through the spatial pyramid matching framework (SPM) and finally classified using a linear SVM. The proposed system has been evaluated using two large-scale tumor cohorts, collected from The Cancer Genome Atlas (TCGA). Experimental results show that the proposed model 1) outperforms systems utilizing sparse coding for unsupervised feature learning (e.g., PSD-SPM [5]); 2) is competitive with systems built upon features with biological prior knowledge (e.g., SMLSPM [4]).

Compressive sensing has recently been applied to electrocardiogram (ECG) acquisition and reconstruction with the aim of lowering energy consumption and sampling rates in wireless body area networks for ambulatory ECG monitoring. However, most current methods only adopt a sparse prior on the ECG wavelet representation. In this paper, we propose to further exploit the wavelet representation structure by incorporating two properties in the formulation of the optimization problem: the exponentially decaying magnitude of the detail coefficients across scales and the accumulation of signal energy in the approximation subband. We derive a weighted ℓ1minimization algorithm, based on a maximum a posteriori (MAP) approach, that leads to a significant reduction in the number of measurements and superior reconstruction performance compared to current CS-based methods with application to wireless ECG systems.

Discovering the intrinsic low-dimensional structure from high-dimensional observation space (e.g., images, videos), in many cases, is critical to successful recognition. However, many existing nonlinear manifold learning (NML) algorithms have quadratic or cubic complexity in the number of data, which makes these algorithms computationally exorbitant in processing real-world large-scale datasets. Randomly selecting a subset of data points is very likely to place NML algorithms at the risk of local optima, leading to poor performance. This paper proposes a novel algorithm called Locality Preserving KSVD (LP-KSVD), which can effectively learn a small number of dictionary atoms as locality-preserving landmark points on the nonlinear manifold. Based on the atoms, the computational complexity of NML algorithms can be greatly reduced while the low-dimensional embedding quality is improved. Experimental results show that LP-KSVD successfully preserves the geometric structure of various nonlinear manifolds and it outperforms state-of-the-art dictionary learning algorithms (MOD, K-SVD and LLC) in our preliminary study on face recognition.

Non-rigid 3D shape recognition is an important and challenging research topic in computer vision and pattern recognition. This paper presents a novel algorithm, called dictionary learning based on supervised locally linear representation (DL-SLLR), for efficient 3D shape recognition using shape descriptors. Specifically, we introduce a novel locality-preservation error term along with a label approximation error term into the objective function. The proposed algorithm optimizes a dictionary for its capability in representation as well as its locality-preservation capability, which thus allows more consistent encoding of similar descriptors compared with sparse coding. In addition, the proposed SLLR coding yields a closed-form solution, compared to many sparse coding algorithms. Experimental results demonstrate that using majority voting, DL-SLLR outperforms D-KSVD and SVM over a newly generated SLI 3D Face Dataset and the SHREC'11 Contest Dataset.

Current nonlinear dimensionality reduction (NLDR) algorithms have quadratic or cubic complexity in the number of data, which limits their ability to process real-world large-scale datasets. Learning over a small set of landmark points can potentially allow much more effective NLDR and make such algorithms scalable to large dataset problems. In this paper, we show that the approximation to an unobservable intrinsic manifold by a few latent points residing on the manifold can be cast in a novel dictionary learning problem over the observation space. This leads to the presented locality constrained dictionary learning (LCDL) algorithm, which effectively learns a compact set of atoms consisting of locality-preserving landmark points on a nonlinear manifold. Experiments comparing state-of-the-art DL algorithms, including K-SVD, LCC and LLC, show that LCDL improves the embedding quality and greatly reduces the complexity of NLDR algorithms.

Chart images visually represent quantitative information. Most of these visual information are represented by graphical symbols and textual descriptions; without access to the Object Model of a graphic it is difficult for viewers to acquire the accurate underlying data. In response we propose the VIEW (Visual Information Extraction Widget), a system that automatically extracts information from raster-format charts to improve accessibility. Taking a chart image as input, the system first segments the image into connected-components and distinguishes them as graphical and textual components. By analyzing the graphical components, the system then identifies the graphic type and further conducts category-specific methods to infer the underlying data. Using the images drawn from the web, we conduct experiments to demonstrate the effectiveness of the proposed system. Based on the extracted information, VIEW generates a general-purpose descriptive data table, leading the production of multi-modal representations under the task-oriented design principle.

In this paper we propose a robust iterative hard thresholding (IHT) algorithm for reconstructing sparse signals in the presence of impulsive noise. To address this problem, we use a Lorentzian cost function instead of the L2cost function employed by the traditional IHT algorithm. The derived algorithm is comparable in computational load to the least squares based IHT. Analysis of the proposed method demonstrates its robustness under heavy-tailed models. Simulations show that the proposed algorithm significantly outperform commonly employed sparse reconstruction techniques in impulsive environments, while providing comparable reconstruction quality in less demanding, light-tailed environments.

Recent works in modified compressed sensing (CS) show that reconstruction of sparse or compressible signals with partially known support yields better results than traditional CS. In this paper, we extend the ideas of these works to modify the iterative hard thresholding (IHT) algorithm to incorporate known support in the recovery process. We present a theoretical analysis that shows that including prior support information relaxes the conditions for stable reconstruction. Numerical results show that the IHT modification improves performance, thereby needing fewer samples to yield an approximate reconstruction.

Compressive sensing (CS) is a new approach for the acquisition and recovery of sparse signals that enables sampling rates significantly below the classical Nyquist rate. Based on the fact that electrocardiogram (ECG) signals can be approximated by a linear combination of a few coefficients taken from a Wavelet basis, we propose a compressed sensing-based approach for ECG signal compression. ECG signals generally show redundancy between adjacent heartbeats due to its quasi-periodic structure. We show that this redundancy implies a high fraction of common support between consecutive heartbeats. The contribution of this paper lies in the use of distributed compressed sensing to exploit the common sup port between samples of jointly sparse adjacent beats. Simulation results suggest that compressed sensing should be considered as a plausible methodology for ECG compression.

The importance of haptic feedback is recognized by an increasing number of researchers in the virtual reality field. Recently, a volume-based haptic feedback approach has emerged. The approach samples the intersection volume between objects through the three axes in 3D space to render an accurate interaction force among objects. This paper presents a method to reduce the complexity of the volume-based force feedback algorithm. The core of the proposed algorithm is sampling the intersection volume between two objects only once rather than three times. For the other two axes, a penetration pair reconstruction algorithm is developed to generate the required information from the sampled result. Experimental results demonstrate that the proposed approach can increase the frame rate of the volumetric haptic feedback algorithm by a factor of over two. The resulting force error is modest compared to the original volume-based haptic feedback. This proposed algorithm may also be applied to accelerate other volume-based applications, e.g. volume based force interaction between colliding deformable objects in virtual reality simulation. Moreover, the algorithm requires no pre-processing, and is thus well suited for simulations where object topology is constantly changing, i.e. cutting, melting or deforming processes.

Information graphics are important visual information in digital media. This paper investigates the accessibility issues associate with the information graphics for visually impaired people. The goal is to provide them with comprehensive numerical information contained in the figures. Towards the goal, we address the practical problems of automatic figure categorization, information extraction and multi-modal presentation scheme. In particular, the system identifies the image class using image processing and machine learning algorithms. With the knowledge of the image class, specific domain features and information are extracted, and then different modalities of presentation are employed based on the need. This paper first proposes the system framework and then focuses on the automated categorization algorithm.

Compressed sensing shows that a sparse or compressible signal can be reconstructed from a few incoherent measurements. Noting that sparse signals can be well modeled by algebraic-tailed impulsive distributions, in this paper, we formulate the sparse recovery problem in a Bayesian framework using algebraic-tailed priors from the generalized Cauchy distribution (GCD) family for the signal coefficients. We develop an iterative reconstruction algorithm from this Bayesian formulation. Simulation results show that the proposed method requires fewer samples than most existing reconstruction methods to recover sparse signals, thereby validating the use of GCD priors for the sparse reconstruction problem.

Recent works in modified compressed sensing (CS) show that reconstruction of sparse or compressible signals with partially known support yields better results than traditional CS. In this paper, we extend the ideas of these works to modify three iterative algorithms to incorporate the known support in the recovery process. The performance and effect of the prior information are studied through simulations. Results show that the modification of iterative algorithms improves their performance, needing fewer samples to yield an approximate reconstruction.

We consider consensus algorithms in their most general setting and provide conditions under which such algorithms are guaranteed to converge, almost surely, to a consensus. Let {A(t), B(t)} ∈ RN×Nbe (possibly) stochastic, nonstationary matrices and {x(t), m(t)} 6 RN×1be state and perturbation vectors, respectively. For any consensus algorithm of the form x(t + 1) = A(t)x(t) + B(t)m(t), we provide conditions under which consensus is achieved almost surely, i.e., Pr-{limt →∞x(t) = c1} -1 for some c ∈ R. Moreover, we show that this general result subsumes recently reported results for specific consensus algorithms classes, including sum-preserving, nonsum-preserving, quantized, and noisy gossip algorithms. Also provided are the e-converging time for any such converging iterative algorithm, i.e., the earliest time at which the vector x(t) is ε close to consensus, and sufficient conditions for convergence in expectation to the average of the initial node measurements. Finally, mean square error bounds of any consensus algorithm of the form discussed above are presented.

Recent results in compressed sensing show that a sparse or compressible signal can be reconstructed from a few incoherent measurements. Compressive sensing systems are not immune to noise, which is always present in practical acquisition systems. In this paper we propose robust methods for sampling and reconstructing sparse signals in the presence of impulsive noise. Analysis of the proposed methods demonstrates their robustness under heavy-tailed models. Simulations show that the proposed methods outperform existing compressed sensing techniques in impulsive environments, while having similar performance in light-tailed environments.

We consider consensus algorithms in their most general setting and provide conditions under which such algorithms are guaranteed to converge, almost surely, to a consensus. Let {A(t),B(t)} isin RNtimesNbe (possibly) random, non-stationary matrices and {x(t),m(t)} isin RNtimes1be state and perturbation vectors, respectively. For any consensus algorithm of the form x(t + 1) = A(t)x(t) + B(t)m(t), we provide conditions under which consensus is achieved almost surely, i.e., Pr {limtrarrinfinx(t) = c1} = 1 for some c isin R. Moreover, we show that this general result subsumes recently reported results for specific consensus algorithms classes, including sum-preserving, non-sum-preserving, quantized and noisy gossip algorithms. Also provided are the e-converging time for any such converging iterative algorithm, i.e., the earliest time at which the vector x(t) is e close to consensus, and sufficient conditions for convergence in expectation to the initial node measurements average.

Infrared sensors are widely utilized on manned and unmanned systems due to their ability to operate during low light conditions as well as their target discrimination capability. Machine vision algorithms that operate on infrared imagery (e.g. target detection, obstacle detection, target tracking) can significantly increase the effectiveness of platforms and the autonomy of unmanned systems. The classification of regions in infrared imagery provides a valuable input to computer vision algorithms. This paper contains an analysis of features for infrared region discrimination, feature dimensionality reduction, and classification for regions of infrared imagery. A variety of features are considered including those based on texture and a physics based model for atmospheric attenuation. An analysis of the optimal feature set and classifier combination is presented. Performance of the classifier on a database of infrared imagery is provided as well as top level contextual techniques to improve classification performance.

The generalized Cauchy distribution (GCD) family has the property that its pdf has closed form for the whole family and also has algebraic tails which makes it suitable to model many impulsive processes in real life. In this paper we propose a robust M-type estimator based on the pdf of the GCD family. Robustness and properties of the new statistics are analyzed and it is noticed that this estimator provide desired characteristics in robust signal processing applications involving non-Gaussian heavy-tailed models. Simulations of the filtering method are performed to evaluate and compare the proposed filtering structure performance to other classic and robust estimators.

A novel method based on empirical mode decomposition (EMD) is introduced in this paper for the detection of affine invariant interest or feature points. The proposed algorithm is a contour based method, where image edges are first detected by utilizing morphological operators followed by an edge thinning process and then the corner or interest points are identified based on the local curvature of the edges. In this work a novel method based on 1-D EMD is formulated to select good discriminative interest points from the edges. The proposed method is compared with four existing approaches that yield good results. The performance is evaluated by employing a criteria known as repeatability rate, which evaluates the geometric stability of an interest point detector under different transformations. The results prove the efficacy and superiority of the proposed technique over other schemes in terms of detecting more true corner points.

This paper addresses text detection utilizing weighted discrete cosine transform (DCT) coefficients as a discrimination statistic. The sum of absolute value of DCT coefficients in each block is considered as an energy statistic for detection. Linear discriminant analysis (LDA) is conducted to calculate the optimal threshold. Different types of weights are employed to strengthen discrimination statistic, which include uniform, binary, linear and quadratic weights. The evaluation of weighted algorithms is conducted in the receiver operating characteristics (ROC) space. The ROC curves show that there is a tradeoff between true positive rate (TPR) and false positive rate (FPR) for all weight configurations. In terms of maximizing the separation between two distributions, experimental results show that the quadratic weighted energy achieves the best recall and precision.

This paper presents an algorithm for extracting and classifying two-dimensional motion in an image sequence based on trajectories. Each gesture signal is represented as a time series in a Principal Component Analysis (PCA) reduced dimensional space. A class of Support Vector Machine (SVM) applicable to sequential-pattern recognition is employed for classification by incorporating a hybrid distance measure into the kernel function that accounts for both the hand shape and movement. The performance of the proposed method is evaluated in continuous tactile hand gesture streams recognition experiments. Results are presented for 9 different gestures performed by 25 subjects at a variety of time scales. Experimental results show that the proposed approach yields high recognition rate for hand gesture motion patterns.

We consider decentralized estimation of a noise- corrupted deterministic signal in a bandwidth-constrained sensor network communicating through an insecure medium. Each sensor collects a noise-corrupted version, performs a local quantization, and transmits a 1-bit message to an ally fusion center through a wireless medium where the sensor outputs are vulnerable to unauthorized observation from enemy/third-party fusion centers. In this paper, we introduce an encrypted wireless sensor network (eWSN) concept where stochastic enciphers operating on binary sensor outputs are introduced to disguise the sensor outputs, creating an eWSN scheme. Noting that the plaintext (original) and ciphertext (disguised) messages are constrained to a single bit due to bandwidth constraints, we consider a binary channel-like scheme to probabilistically encipher (i.e., flip) the sensor outputs. We first consider a symmetric key encryption case where the “0” and “1” enciphering probabilities are equal. The key is represented by the bit enciphering probability. Specifically, we derive the optimal estimator of the deterministic signal approached from a maximum-likelihood perspective and the Cramer-Rao lower bound for the estimation problem utilizing the key. Furthermore, we analyze the effect of the considered cryptosystem on enemy fusion centers that are unaware of the fact that the WSN is encrypted (i.e., we derive the bias, variance, and mean square error (MSE) of the enemy fusion center). We then extend the cryptosystem to admit unequal enciphering schemes for “0” and “1”, and analyze the estimation problem from both the prospectives of ally (that has access to the enciphering keys) and (third-party) enemy fusion centers. The results show that when designed properly, a significant amount of bias and MSE can be introduced to an enemy fusion center with the cost to the ally fusion center being a marginal increase [factor of$(1-\Omega _{1}-\Omega _{0})^{-2}$, where$1-\Omega _{j}$,$j=0,1$is the “$j$” enciphering probability] in the estimation variance (compared to the variance of a fusion center estimate operating in a vulnerable WSN).

Recently proposed decentralized, distributed estimation and power scheduling methods for wireless sensor networks (WSNs) do not consider errors occurring during the transmission of binary observations from the sensors to fusion center. In this letter, we extend the decentralized estimation model to the case in which imperfect transmission channels are considered. The proposed estimators, which operate on additive channel noise corrupted versions of quantized noisy sensor observations, are approached from a maximum likelihood (ML) perspective. Complicating this approach is the fact that the noise distribution is rarely fully known to the fusion center. Here we assume the distribution is known but not the defining parameters, e.g., variance. The resulting incomplete data estimation problem is approached from a expectation-maximization (EM) perspective. The critical initialization and convergence aspects of the EM algorithm are investigated. Furthermore, the estimation of the source parameter is extended to the blind case where both the channel and sensor noise parameters are unknown. Finally, numerical experiments are provided to show the effectiveness of the proposed estimators.

This paper extends the homogeneous wireless sensor network (WSN) model to possess (possibly) varying levels of noise that represent environment clusters in a WSN. The estimation of the source parameter is approached from maximum likelihood (ML) perspective. The Cramer Rao lower bound (CRLB) for any unbiased estimator operating in a clustered WSN environment is derived. Noting that the ML estimate cannot be found in closed-from, we resort to a numerical search. Although numerically determined, the ML estimate is guaranteed to converge to the optimal solution since it is shown here that the log-likelihood function is concave. Also considered is the estimation of a random parameter with a priori information, which is approached from maximum a posteriori (MAP) perspective. Finally, the proposed MAP and ML optimal are validated and compared to theoretical bounds with illustrative numerical examples.

The linear, median, myriad filtering structures are statistically related to the maximum likelihood (ML) estimates of location under Gaussian, Laplacian, and Cauchy statistics, respectively In this paper, we propose a filtering structure based on the ML estimate of the constructed and so-called meridian statistics. Analysis and simulations presented here indicate that the proposed filtering structure exhibits characteristics more robust than that of mean, median and myriad filtering structures.

Recently proposed decentralized estimation methods do no consider errors occurring during the transmission of binary observations from the sensors to fusion center. In this paper, we extend the decentralized estimation model to the case where imperfect transmission channels are considered. The proposed estimator, which operates on additive channel noise corrupted versions of quantized noisy sensor observations, is approached from maximum likelihood (ML) perspective. The ML estimate unfortunately has no closed-form solution. We analyze the log-likelihood function showing that it is log-concave, thereby indicating that numerical methods, such as Newtons algorithm, can be utilized to obtain the optimal solution. A suboptimal mean estimator that requires minimal information about the channel and sensing environment is also proposed. Simulation results evaluating the variances of the proposed optimal and suboptimal solutions are provided.

In this paper, we extend the decentralized estimation model to the case in which imperfect transmission channels are considered. The proposed estimators, which operate on additive channel noise corrupted versions of quantized noisy sensor observations, are approached from a maximum likelihood (ML) perspective. Complicating this approach is the fact that the noise distribution is rarely fully known to the fusion center. Here we assume the distribution is known but not the defining parameters, e.g., variance. The resulting incomplete data estimation problem is approached from a expectation-maximization (EM) perspective. The critical initialization and convergence aspects of the EM algorithm are investigated. Furthermore, the estimation of the source parameter is extended to the blind case where both the channel and sensor noise parameters are unknown. Finally, numerical experiments are provided to show the effectiveness of the proposed estimators.

The empirical mode decomposition (EMD) was recently proposed as a new time-frequency analysis tool for nonstationary and nonlinear signals. Although the EMD is able to find the intrinsic modes of the signal and is completely self-adaptive, it does not have any implication on optimality. In some situation, when certain optimality is considered, we need a more flexible signal decomposition and reconstruction scheme. We propose a modified version of the EMD, which enhances the capability of the EMD. The proposed modified EMD algorithm gives the best estimate to a given signal in the minimum mean square error sense. Two different formulations are proposed. The first one utilizes a linear weighting for the intrinsic mode functions (IMF). The second algorithm adopts a bidirectional weighting, namely, it not only uses weighting for IMF modes, but also exploits the correlations between samples in a specific window and carries out filtering in the window. These two new EMD methods extend the capability of the traditional EMD and is well suited for optimal signal recovery. Simulation studies are performed to show the application of the proposed optimal EMD algorithms to denoising problem.

This paper describes a method of image segmentation using recursive splitting and merging based on texture similarity measures. This technique addresses the problem of segmenting image regions of varying texture with limited intensity edges. The proposed technique provides a framework for texture based image segmentation that is shown to be applicable across a wide variety of image content. The primary motivation for this work is the segmentation of infrared images. Infrared imagery is characterized by narrow histograms corresponding to ambient scene temperature. Results illustrate that using texture signatures for infrared imagery yields enhanced segmentation performance over luminance features. Additional benefit for infrared imagery and better generality to other images types is obtained when luminance and texture are both applied to the segmentation criteria. A method for the quantitative comparison of segmentation results is presented and benchmarks are provided against several recent segmentation algorithms.

We propose a computationally efficient super-resolution (SR) algorithm to produce high-resolution video from low-resolution (LR) video using partition-based weighted sum (PWS) filters. First, subpixel motion parameters are estimated from the LR video frames. These are used to position the observed LR pixels into a high-resolution (HR) grid. Finally, PWS filters are employed to simultaneously perform nonuniform interpolation (to fully populate the HR grid) and perform deconvolution of the system point spread function. The PWS filters operate with a moving window. At each window location, the output is formed using a weighted sum of the present pixels within the window. The weights are selected from a filter bank based on the configuration of missing pixels in the window and the intensity structure of the present pixels. We present an algorithm for applying the PWS SR filters to video that is computationally efficient and suitable for parallel implementation. A number of experimental results are presented to demonstrate the efficacy of the proposed algorithm in comparison to several previously published methods. A detailed computational analysis of the partition-based SR filters is also presented

Image sequence prediction is widely used in image compression and transmission schemes such as differential pulse code modulation. In traditional predictive coding, linear predictors are usually adopted for simplicity. The nonlinear Volterra predictor can be employed as an alternative to linear predictors to compensate for the nonstationary and non-Gaussian nature of image sequences. Although the Volterra predictor avoids the smoothing effects introduced by linear predictors, it generally amplifies noise contamination present in the images. In this letter, we propose a nonlinear polynomial weighted median (PWM) predictor for image sequence. The proposed PWM predictor is more robust to noise, while still retaining the information of higher order statistics of pixel values. Experimental results illustrate that the PWM predictor yields good results in both high and low motion video. It is especially suitable for high motion sequence in noisy case. The proposed scheme can be incorporated in new predictive coding systems.

Speckle is a multiplicative noise that degrades ultrasound images. Recent advancements in ultrasound instrumentation and portable ultrasound devices necessitate the need for more robust despeckling techniques, for both routine clinical practice and teleconsultation. Methods previously proposed for speckle reduction suffer from two major limitations: 1) noise attenuation is not sufficient, especially in the smooth and background areas; 2) existing methods do not sufficiently preserve or enhance edges--they only inhibit smoothing near edges. In this paper, we propose a novel technique that is capable of reducing the speckle more effectively than previous methods and jointly enhancing the edge information, rather than just inhibiting smoothing. The proposed method utilizes the Rayleigh distribution to model the speckle and adopts the robust maximum-likelihood estimation approach. The resulting estimator is statistically analyzed through first and second moment derivations. A tuning parameter that naturally evolves in the estimation equation is analyzed, and an adaptive method utilizing the instantaneous coefficient of variation is proposed to adjust this parameter. To further tailor performance, a weighted version of the proposed estimator is introduced to exploit varying statistics of input samples. Finally, the proposed method is evaluated and compared to well-accepted methods through simulations utilizing synthetic and real ultrasound data.

In this paper, we propose a novel speckle filtering algorithm based on the maximum likelihood estimate of speckle statistics. The homomorphism approach is utilized to modify the multiplicative noise problem to an additive one. The statistics of the altered noise are derived. The problem is then converted to the statistical M-estimation. Properties of the proposed filtering structure are investigated. Finally, simulations indicating the advantages of the proposed filtering are presented.

Image sequence prediction is widely used in image compression and transmission schemes such as differential pulse code modulation (DPCM). In traditional predictive coding, linear predictors are usually adopted to exploit the inherent redundancy and correlation between neighboring pixels. However, due to the nonstationary and non-Gaussian nature of image sequences, linear predictors are not often very effective. As an alternative, Volterra predictor is able to compensate for the smoothing effects introduced by linear predictor. However, it suffers from noise that may be attributed to quantization errors or image acquisition devices. In this paper, we propose a novel nonlinear polynomial weighted median (PWM) predictor for image sequence prediction. The proposed PWM predictor is more robust to noise, while still retaining the information of higher-order statistics of pixel values. Experimental results illustrate that the PWM predictor yields better results than other predictors especially in noisy case. The proposed scheme can be incorporated in new predictive coding systems.

Watershed algorithms were originally developed as gray-level dedicated transformations. However, well known scalar watershed algorithms can be easily adapted to include multispectral information, producing color segmentations in agreement with the human perception. The multispectral watershed algorithm introduced in this paper includes the use of a vector gradient for calculating the gradient magnitude image (GMI) and the choice of a uniform color space, such as the L*a*6*, for adapting the dissimilarity measures utilized in the merging stage. Experimental results are presented showing the advantages of the proposed adaptation method as well as advantages of combined homogeneity-edge integrity region merging criterion.

Methods to automatically convert graphics into tactile representations have been recently investigated, creating either raised-line or relief images. In particular, we briefly review one raised-line method where important features are emphasized. This paper focuses primarily on the effects of such emphasis and on comparing both raised-line and relief methods (produced by a Tiger Braille printer) through psychophysical experiments including discrimination, identification, and comprehension (involving 14 sighted and 6 blind subjects). Results show that raised-line pictures outperform relief images in all three tasks. Also, emphasizing important features significantly increases comprehension of tactile graphics. Raised-line images, however, disregard intensity/colour information unlike relief pictures. A technique incorporating intensity into raised-line graphics is also presented. Further experiments show the validity of this technique.

Quadratic Volterra filters are effective in image sharpening applications. The linear combination of polynomial terms, however, yields poor performance in noisy environments. Weighted median (WM) filters, in contrast, are well known for their outlier suppression and detail preservation properties. The WM sample selection methodology is naturally extended to the quadratic sample case, yielding a filter structure referred to as quadratic weighted median (QWM) that exploits the higher order statistics of the observed samples while simultaneously being robust to outliers arising in the higher order statistics of environment noise. Through statistical analysis of higher order samples, it is shown that, although the parent Gaussian distribution is light tailed, the higher order terms exhibit heavy-tailed distributions. The optimal combination of terms contributing to a quadratic system, i.e., cross and square, is approached from a maximum likelihood perspective which yields the WM processing of these terms. The proposed QWM filter structure is analyzed through determination of the output variance and breakdown probability. The studies show that the QWM exhibits lower variance and breakdown probability indicating the robustness of the proposed structure. The performance of the QWM filter is tested on constant regions, edges and real images, and compared to its weighted-sum dual, the quadratic Volterra filter. The simulation results show that the proposed method simultaneously suppresses the noise and enhances image details. Compared with the quadratic Volterra sharpener, the QWM filter exhibits superior qualitative and quantitative performance in noisy image sharpening.

The spatial and rank (SR) orderings of samples play a critical role in most signal processing algorithms. The recently introduced fuzzy ordering theory generalizes traditional, or crisp, SR ordering concepts and defines the fuzzy (spatial) samples, fuzzy order statistics, fuzzy spatial indexes, and fuzzy ranks. Here, we introduce a more general concept, the fuzzy transformation (FZT), which refers to the mapping of the crisp samples, order statistics, and SR ordering indexes to their fuzzy counterparts. We establish the element invariant and order invariant properties of the FZT. These properties indicate that fuzzy spatial samples and fuzzy order statistics constitute the same set and, under commonly satisfied membership function conditions, the sample rank order is preserved by the FZT. The FZT also possesses clustering and symmetry properties, which are established through analysis of the distributions and expectations of fuzzy samples and fuzzy order statistics. These properties indicate that the FZT incorporates sample diversity into the ordering operation, which can be utilized in the generalization of conventional filters. Here, we establish the fuzzy weighted median (FWM), fuzzy lower-upper-middle (FLUM), and fuzzy identity filters as generalizations of their crisp counterparts. The advantage of the fuzzy generalizations is illustrated in the applications of DCT coded image deblocking, impulse removal, and noisy image sharpening.

The rank information of samples is widely utilized in nonlinear signal processing algorithms. Recently developed fuzzy transformation theory introduces the concept of fuzzy ranks, which incorporates sample spread (or sample diversity) information into the sample ranking framework. Thus, the fuzzy rank reflects a sample's rank, as well as its similarity to the other sample (namely, joint rank order and spread), and can be utilized to improve the performance of the conventional rank-order-based filters. In this paper, the well-known lower-upper-middle (LUM) filters are generalized utilizing the fuzzy ranks, yielding the class of fuzzy rank LUM (F-LUM) filters. Statistical and deterministic properties of the F-LUM filters are derived, showing that the F-LUM smoothers have similar impulsive noise removal capability to the LUM smoothers, while preserving the image details better. The F-LUM sharpeners are capable of enhancing strong edges while simultaneously preserving small variations. The performance of the F-LUM filters are evaluated for the problems of image impulsive noise removal, sharpening and edge-detection preprocessing. The experimental results show that the F-LUM smoothers can achieve a better tradeoff between noise removal and detail preservation than the LUM smoothers. The F-LUM sharpeners are capable of sharpening the image edges without amplifying the noise or distorting the fine details. The joint smoothing and sharpening operation of the general F-LUM filters also showed superiority in edge detection preprocessing application. In conclusion, the simplicity and versatility of the F-LUM filters and their advantages over the conventional LUM filters are desirable in many practical applications. This also shows that utilizing fuzzy ranks in filter generalization is a promising methodology.

Partition-based Weighted Sum (P-WS) filtering is an effective method for processing nonstationary signals, especially those with regularly occurring structures, such as images. P-WS filters were originally formulated as Hard-partition Weighted Sum (HP-WS) filters and were successfully applied to image denoising. This formulation relied on intuitive arguments to generate the filter class. Here we present a statistical analysis that justifies the use of weighted sum filters after observation space partitioning. Unfortunately, the HP-WS filters are nondifferentiable and an analytical solution for their global optimization is therefore difficult to obtain. A two-stage suboptimal training procedure has been reported in the literature, but prior to this research no evaluation on the optimality of this approach has been reported. Here, a Genetic Algorithm (GA) HP-WS optimization procedure is developed that, in simulations, shows that the simpler two-stage training procedure yields near optimal results. Also developed in this paper are Soft-partition Weighted Sum (SP-WS) filters. The SP-WS filters utilize soft, or fuzzy, partitions that yield a differentiable filtering operation, enabling the development of gradient-based optimization procedures. Image denoising simulation results are presented comparing HP-WS and SP-WS filters, their optimization procedures, and wavelet-based image denoising. These results show that P-WS filters, in general, outperform traditional and wavelet-based image filters, and SP-WS filters utilizing soft partitioning not only allow for simple optimization, but also yields improved performance.

In this study, the accuracy of an optical mouse, optical trackball, isotonic joystick, and a FingerWorks MultiTouch Surface (MTS) are compared for users suffering from Parkinsonian tremor and essential tremor. Using a data acquisition program, WinFitts, created at the University of Oregon's HCI Lab, data collected from five subjects with Parkinsonian tremor, five with essential tremor, and eleven with no tremor is analyzed and compared. Both temporal and spatial analyses are obtained from all of the subject data. The time-based measures of performance for each device include the uses of Fitts' law and the Proximity Movement Time, while the spatially-based measures include the use of the Deviation Accuracy and the Click Histogram. A statistical analysis is performed using a t-test to show the differences between the resulting means of some of the measures. By using the MUSIC spectral estimation technique, an analysis of the frequency and the amplitude of the tremor showed how certain devices performed in hand tremor suppression.

This paper presents a method to convert vector graphics into tactile representations for the blind. Generating tactile pictures from vector graphics is an important effort to bring more accessibility to the WWW as well as other means of communications since vector graphics are an increasing trend in web based graphics. Prior research has investigated methods that extracts object boundaries from images to produce raised-line tactile pictures. The proposed method extends this idea for vector graphics, producing tactile pictures where important outlines are emphasized. Important outlines are determined by using the hierarchical structure of a vector graphic. A Braille printer is used where raised dots are embossed for the outlining boundaries. Important and detail boundaries are embossed with dots of larger and smaller height, respectively, while all other regions contain no raised dots. Results testing a person's ability to discriminate, identify, and comprehend tactile pictures shows the proposed methods' advantage over two other methods.

Polynomial filter theory methods are able to approximate important classes of nonlinear systems. In the impulsive noise environments, however, outliers are amplified by the polynomial structure, yielding poor performance. Median filters are well-known for their impulsive noise suppression and fine-detail preservation. In this paper, we extend the weighted median (WM) filtering to a polynomial class, polynomial weighted median (PWM) filters which exploit higher order statistics while being robust to outliers. The PWM filter class is motivated by a study of the statistics of cross and square terms. The breakdown probability for the PWM is given and weight optimization is presented. Finally, the effectiveness of the PWM filter is shown through simulations.

Wavelet transforms are extensively used for image denoising and compression problems. The sparse property is the major reason for the effectiveness of the nonlinear operation such as thresholding of wavelet-transformed coefficients. Applying the thresholds to all coefficients uniformly, however, produces oversmoothing results on edges, or undersmoothing on uniform regions. Using statistical parameterized models often produces some artifacts because of overparameterizing. Recently, adaptive wavelet thresholding utilizing the correlation of space and adjacent scale has been introduced. This paper introduces a related, but more direct, technique of adaptively processing wavelet coefficients based on partitioning of the coefficient space. In the wavelet domain, the coefficient space is partitioned by a vector quantization method and the mask functions are used to obtain the denoised wavelet coefficients. Simulations show that the proposed technique yields a superior performance compared with current wavelet denoising methods.

The previously proposed partition-based weighted sum (PWS) filters combine vector quantization (VQ) and linear finite impulse response (FIR) Wiener filter concepts. By partitioning the observation space and applying a tuned Wiener filter to each partition, the PWS is spatially adaptive and has been shown to perform well in noise reduction applications. In this paper, we propose the subspace PWS (SPWS) filter and evaluate the efficacy of the SPWS filter applied to the image deconvolution problem. In the SPWS filter, we project the observation vectors into a subspace using principal component analysis (PCA) for partitioning. This subspace projection can dramatically reduce the computational burden associated with the large window size PWS filters that are needed for effective image deconvolution. In some cases, performance is also enhanced due to improved partitioning.

The paper presents a new method using fuzzy filtering to remove the coding artifacts in compressed video. The method takes the interlaced video format into consideration and processes each field separately. For deblocking, a 1D fuzzy filter with different window sizes is used to remove the horizontal and vertical blocking artifacts respectively. For deringing, each 8/spl times/8 block in a field is first classified into one of the four categories, i.e., strong edge, weak edge, texture and smooth blocks. According to each block's type and the neighboring block's type, the spread parameter of a 2D fuzzy filter is adaptively decided and the filter is applied To speed up the process, the fuzzy filter weights are generated using a piecewise linear membership function instead of the conventional Gaussian function. Experimental results show that the proposed method has better detail preservation and lower computational costs than our previous method. It achieves comparable deblocking and superior deringing performance to the MPEG-4 standard method at similar computation costs.

The MUSIC method represents a class of super-resolution methods for frequency estimation. However, it has poor performance in impulsive noise environments due to the presence of outliers. A more robust method called trimmed correlation based-MUSIC (TR-MUSIC) method is proposed in this paper. Through a trimming operation, outliers in the samples participating in the correlation calculation are discarded. The amount of trimming is determined by the Mahalanobis distance in which robust estimates of location and scale are utilized. Frequency estimation results from the eigendecomposition of the trimmed correlation matrix. Corroborating simulations are presented to show the robustness and performance improvement of the proposed method.

Multi-touch technology provides a successful gesture based human computer interface. The contact and gesture recognition algorithms of this interface are based on full hand function and, therefore, are not accessible to many people with physical disability. In this paper, we design a set of command-like gestures for users with limited range and function in their digits and wrist. Trajectory and angle features are extracted from these gestures and passed to a recurrent neural network for recognition. Experiments are performed to test the feasibility of the gesture recognition system and determine the effect of network topology on the gesture recognition rate. These results show that the proposed method can successfully recognize those designed gestures for disabilities.

The previously proposed partition-based weighted sum (PWS) filters combine vector quantization (VQ) and linear finite impulse response (FIR) Wiener filtering concepts. By partitioning the observation space and applying a tuned Wiener filter to each partition, the PWS is spatially adaptive and has been shown to perform well in noise reduction applications. In this letter, we propose the subspace PWS (SPWS) filter and evaluate the efficacy of the SPWS filter in image deconvolution and noise reduction applications. In the SPWS filter, we project the observation vectors into a subspace using principal component analysis (PCA), or other methods, prior to partitioning. This subspace projection can dramatically reduce the computational burden associated with partitioning, especially for large window sizes. In some cases, performance is also enhanced due to improved partitioning.

The paper focuses on the queue size estimation problem in random early detection (RED) gateways. Queue size estimation plays a critical role in gateways' packet dropping/marking decisions. Conventional RED gateways use exponentially weighted moving averages (EWMA) to estimate the queue size. These IIR filters require very small weights in order to avoid nonlinear instabilities and accommodate transient congestion. Small weights, however, lead to failure of gateways to track rapid queue size depletion closely and thus causes link under utilization. We use adaptive weighted median filters for queue size estimation and study the corresponding queue dynamics. Simulation results show that the proposed algorithm provides better stability in queue dynamics, greater network power, less global synchronization, and a fairer treatment to bursty traffic than the RED algorithm.

Fuzzy ranks are a component of fuzzy ordering theory developed to incorporate sample spread information into sample ranking. In this paper, the well-known lower-upper-middle (LUM) filters are generalized by utilizing the fuzzy ranks, which is called fuzzy LUM (F-LUM) filters. We develop the statistical and deterministic properties of the F-LUM filters and present their performance in image denoising and enhancing applications. As shown by the experimental results, the F-LUM filters inherit the simplicity and versatility of the LUM filters such that they can perform signal smoothing, sharpening and outlier rejection using a single filter structure, while having great improvement in detail preservation and robustness to noise.

In this paper, we analyze several previous optimization approaches for weighted vector median (WVM) filters and show their deficiencies. We then propose two fast adaptive WVM optimization algorithms. Proposed algorithm I computes the optimal weight changes at each iteration, and updates weights accordingly. Proposed algorithm II extends the results from weighted median optimization to the vector case by a generalization of an error metric. Both algorithms are fast and stable, and perform well under a wide variety of circumstances.

The work presents a new adaptive approach for blocking and ringing artifact reduction. In order to avoid smearing of the image details, the proposed method first performs visual artifact detection and then applies adaptive filtering to the corrupted blocks. Both visual artifact detection and filtering are guided by an edge map which is constructed based on local features. A fuzzy identity filter is used for image de-ringing. Since it possesses a good edge preserving property and the filtering operation is applied to the edge blocks only (smooth and textured blocks are unaltered), the proposed method shows great effectiveness of both artifact reduction and detail preservation. Experiments demonstrate better results compared with other methods.

This letter presents a new retransmission timeout (RTO) algorithm based on recursive weighted median (RWM) filters for the transmission control protocol (TCP). The RTO algorithm utilized in current TCP implementations is Jacobson's algorithm, which is based on recursive linear filtering. While linear filters are adequate for estimation in Gaussian signal environments, the round trip time (RTT) signals filtered to determine the RTOs are often impulsive. Thus, Jacobson's algorithm is not effective in many cases. The proposed algorithm employs RWM filters that yield improved performance when operating on RTT signals with heavy tailed statistics. Simulation results show that the proposed method yields significantly tighter RTT bounds than Jacobson's method over Internet traffic with heavy tailed statistics.

This paper proposes a novel approach for smoothing surfaces represented by triangular meshes. The proposed method is a two-step procedure: surface normal smoothing through fuzzy vector median (FVM) filtering followed by integration of surface normals for vertex position update based on the least square error (LSE) criteria. Median and Order Statistic-based filters are extensively used in signal processing, especially image processing, due to their ability to reject outliers and preserve features such as edges and monotonic regions. More recently, fuzzy ordering theory has been introduced to allow averaging among similarly valued samples. Fuzzy ordering theory leads naturally to the fuzzy median, which yields improved noise smoothing over traditional crisp median filters. This paper extends the fuzzy ordering concept to vector-based data and introduces the fuzzy vector median filter. The application of FVM filters to surface normal smoothing yields improved results over previously introduced normal smoothing algorithms. The improved filtering results, coupled with LSE vertex position update, produces surface smoothing that minimizes the effects of noise while simultaneously preserving detail features. The proposed method is simple to implement and relatively fast. Simulation results are presented showing the performance of the proposed method and its advantages over commonly used surface smoothing algorithms. Additionally, optimization procedures for FVM filters are derived and evaluated.

In this paper, we introduce the basic concepts of the fuzzy transformation (FZT) theory and their applications in image denoising. FZT is a data-adaptive nonlinear transformation, where the transformation matrix is a dynamic real-valued spatial-rank (SR) matrix (i.e., fuzzy SR matrix) generated by using an order invariant membership function. The clustering property is the essential property of FZT, and is exploited (through the re suiting fuzzy samples and fuzzy ranks) in two image denoising applications, namely, DCT coded image deblocking and impulsive noise elimination. Our proposed methods demonstrate superior performance than the conventional methods in all the simulations, showing that FZT is a powerful tool for image processing.

Images are nonstationary in nature. Hard-partition-based weighted sum (HWS) filters give good results in handling both flat regions and details in application such as image enhancement. But due to their non-differentiable filtering operation, their global optimization is difficult to achieve. In this paper, we introduced a new type of soft-partition-based weighted sum (SWS) filters, whose filter operations are differentiable, allowing a steepest decent adaptive optimization to be applied. It is shown that SWS filters provide better performance than HWS filters in an image enhancement application. SWS filters are, therefore, an effective class of nonlinear filters for image enhancement.

In this paper, the concept of fuzzy transformation is developed based on fuzzy ordering theory. The clustering property of the fuzzy transformation is investigated by deriving the distribution function of the fuzzy samples and is exploited to address the DCT coded image deblocking problem. In this application, the fuzzy transformation is actually a highly data-adaptive smoothing and sharpening process. The spread parameter of the fuzzy transformation controls the smoothing level and can be optimized to fit images at different compression levels. The simulation results show that the fuzzy transformation can effectively remove various blocking artifacts and restore the original image monotone surfaces as well as edges with low computational complexity.

This paper proposes a novel approach for smoothing surfaces represented by triangular meshes. The proposed method is a two-step procedure: surface normal smoothing through fuzzy vector median (FVM) filtering followed by vertex position update based on the least square error criteria. This paper applies the fuzzy ordering theory to the vector based data and introduces the fuzzy vector median. The application of FVM filters to surface normal smoothing yields improved results over previously introduced normal smoothing algorithms. The improved filtering results, coupled with LSE vertex position update, produces surface smoothing that minimizes the effects of noise while simultaneously preserving detail features. Simulation results are presented showing the performance of the proposed method and its advantages over commonly used surface smoothing algorithms. Additionally, optimization procedures for FVM are derived and evaluated.

Permutation filters are a broad class of nonlinear selection filters that utilize the complete spatial and rank order information of observation samples. This use of joint spatial-rank information has proven useful in numerous applications. The application of permutation filters, however, is limited by the factorial growth in the number of spatial-rank orderings. Although M-permutation filters have been developed to address the growth in orderings, their a priori uniform selection of samples is not appropriate in most cases. Permutation filter implementations based on acyclic connected graphs provide a more general approach that allows the level of ordering information utilized to automatically adjust to the problem at hand. In addition to developing and analyzing graph implementations of permutation filters this paper presents a LNE based optimization of the graph structure and filter operation. Simulation results illustrating the performance of the optimization technique and the advantages of the graph implementation are presented.

Combining spatial and rank (SR) order information into filtering methods has been exploited in recent nonlinear filtering algorithms. Fuzzy ordering theory incorporates sample spread information into the SR ranking framework and develops the concepts of fuzzy ranking and fuzzy order statistic. In this paper, fuzzy concepts are utilized to generalize weighted median filters and optimization of the filter parameters is discussed. Simulation results show that the proposed fuzzy weight median (FWM) filter significantly outperforms its crisp counterpart and existing FWM techniques in the literature in image denoising and image zooming.

This paper develops nonlinear multiresolution techniques for scientific visualization utilizing haptic methods. The visualization of data is critical to many areas of scientific pursuit. Scientific visualization is generally accomplished through computer graphic techniques. Recent advances in haptic technologies allow visual techniques to be augmented with haptic methods. The kinesthetic feedback provided through haptic techniques provides a second modality for visualization and allows for active exploration. Moreover, haptic methods can be utilized by individuals with visual impairments. Haptic representations of large data sets, however, can be confusing to a user, especially if a visual representation is not available or cannot be used. This paper develops a multiresolution data decomposition method based on the affine median filter. This results in a hybrid structure that can be tuned to yield a decomposition that varies from a linear wavelet decomposition to that produced by the median filter. The performance of this hybrid structure is analyzed utilizing deterministic signals and statistically in the frequency domain. This analysis and qualitative and quantitative implementation results show that the affine median decomposition has advantages over previously proposed methods. In addition to multiresolution decomposition development, analysis, and results, haptic implementation methods are presented.

A new image segmentation method is proposed for the automatic conversion of images from visual to tactile form. The proposed method utilizes a watershed-based algorithm for obtaining the initial segmentation. A new joint region merging criterion is developed to reduce the number of initial regions in a more appropriate way. The joint criterion combines, in a weighted fashion, criteria based on region homogeneity and edge integrity, which have each been applied marginally with some relative success. However, each criteria has shown some significant drawbacks when applied independently. The new criterion takes joint advantage of these two methods, giving a final segmentation that is more visually appropriate. Experimental results are presented showing the advantages of the new merging criterion. Also, the proposed method is compared with multiresolution (MR) edge detection techniques for tactile imaging applications. Additional results are included showing the advantages of the segmentation procedure over MR edge detection techniques.

Permutation filters are a broad class of nonlinear selection filters that utilize the complete spatial and rank order information of observation samples. The application of permutation filters is limited by the factorial growth in the number of spatial-rank orderings. Although M-permutation filters, which consider the spatial-rank ordering of a fixed sample subset, have been developed to address the growth in orderings, this method requires an a priori selection of samples that is uniformly applied. This uniform application is not appropriate in most cases. We develop a more general multiresolution approach based on decision trees. This implementation allows the level of ordering information utilized to automatically adjust to the problem at hand. Optimization procedures are developed for this method and simulation results illustrating its advantages are presented.

A new joint region merging criterion for watershed-based image segmentation is proposed. The method combines, in a weighted fashion, criteria based on region homogeneity and edge integrity, which have each been applied marginally with some relative success. However, those criteria have shown some significant drawbacks when applied separately. The new criterion takes joint advantage of these two methods, giving a final segmentation that is more visually appropriate. Additionally, a new nonlinear pre and post-filtering procedure is proposed for reducing the oversegmentation of the watershed algorithm. Experimental results are presented showing the advantages of the new merging criterion and the pre/post-filtering method.

This paper examines the application of extended permutation rank selection (EPRS) filters in image interpolation. EPRS filters are constrained to output an order statistic based on N observation samples and K statistics that are functions of the observation samples. By including the sample mean as the sole statistic when K=1, EPRS filters were shown to have superior edge-enhancing properties. By letting K=0, we can define a subset class known as rank conditioned rank selection (RCRS) filters. We show that these filters can also be applied to the issue of image interpolation. In this case, by extending the observation space by including more original samples and by adding additional linear statistics, EPRS filters produce results superior to traditional methods in the application of image interpolation.

In this work, we develop the concept of partitioning the observation space to build a general class of filters referred to as partition-based weighted sum (PWS) filters. In the general framework, each observation vector is mapped to one of M partitions comprising the observation space, and each partition has an associated filtering function. We focus on partitioning the observation space utilizing vector quantization and restrict the filtering function within each partition to be linear. In this formulation, a weighted sum of the observation samples forms the estimate, where the weights are allowed to be unique within each partition. The partitions are selected and weights tuned by training on a representative set of data. It is shown that the proposed data adaptive processing allows for greater detail preservation when encountering nonstationarities in the data and yields superior results compared to several previously defined filters. Optimization of the PWS filters is addressed and experimental results are provided illustrating the performance of PWS filters in the restoration of images corrupted by Gaussian noise.

Edge detection is an important low-level vision problem. Most edge detection methods operate on an image at a single resolution and output a binary edge map. Edges within an image, however, generally occur at various resolutions, or scales, and represent transitions of different degrees, or gradient levels. Thus, single resolution edge detection methods that output binary edge maps do not always yield satisfactory results. This paper develops a multiresolution edge detection method that utilizes a multirate wavelet decomposition to generate a series of images with progressively lower edge resolution. Edges are then recursively extracted to form series of edge maps where the output is not restricted to be binary and is set to reflect the gradient level at each edge point. The series of edge maps is restricted to form a stacking edge map pyramid. In this formulation, the base (lowest level) edge map contains edges at all scales while edge pruning, based on edge scale, is performed at subsequent levels. This approach is shown to have advantages over previously defined multiresolution edge detection methods. Results are presented using natural seen images. These results are motivated by tactile imaging, an important problem in making visual information accessible to blind individuals.

The rank ordering of samples is widely used in robust statistics and robust signal processing. Advances in these areas have focused on utilizing joint time-rank (TR) information. The TR information utilized to date is that resulting from a binary, or crisp, relation between the marginal time and rank ordering of samples. This crisp relation, while powerful, contains no information on sample values or spread. This paper generalizes the TR relation through fuzzy set theory. This generalization includes information on sample spread and leads to the concepts of fuzzy TR relations, fuzzy time and rank ordered samples, and fuzzy time and rank indices. These concepts are developed and analyzed through the derivation of fundamental properties. It is shown that the fuzzy TR relations, samples, and indices contain their crisp (standard) counterparts as special cases. These fuzzy generalizations constitute powerful tools that can be exploited in the design of signal processing algorithms.

We introduce a novel, data-adaptive, and robust filtering framework: affine order statistic filters. Affine order statistics relate classical order statistics to observations in their natural order and thus inherently yield a meaningful data representation. Affine order statistic filters exploit this notion to adaptively process nonstationary signals. Affine order statistic filters overcome many of the limitations associated with traditional order statistic filters, in particular: filters in this class are parsimonious in the number of filter coefficients, they are statistically efficient in a wide range of signal statistics, and they admit real-valued filter weights leading to a wide-range of filtering characteristics. The class of affine order statistic filters contains two families: the weighted order statistic (WOS) affine fitter class whose structure can adapt, according to the observed data, from an FIR linear filter to a WOS filter, and the FIR affine filter class whose structure can adapt from an L-filter to an FIR-filter. We introduce the median affine filter and the center affine filter as representatives of each class, and show their performance in two applications where the signals are nonstationary in nature.

This paper investigates compression of sign-language video sequences for transmission over low bitrate channels. The motion of hands and arms inherent in sign language requires high temporal resolution and greater compression for band-limited channels. Most existing compression schemes treat the entire image uniformly. However, the background is much less critical than the foreground for sign language communication. A new algorithm is developed that gives priority to the foreground and sacrifices the quality of the background to achieve higher compression rates and higher foreground quality without compromising the frame rate. The algorithm is extended to use a dynamic region of interest defined by face and hand location. Results show that the frame rate and quality of the region of interest are higher than with uniform algorithms.

Speech pitch detection remains a fundamental problem due to importance in numerous aspects of speech processing.Current pitch detectors focus on determining the Glottal Closure Instant (GCI).Accurate GCI measures can be obtained from the Dierentiated Electroglottograph (DEGG) signal.Unfortunately, DEGG signals are not available in most practical applications.A n o v el method of pitch detection is proposed here based on the nonlinear estimation of DEGG signals from the acoustic speech w a v eform.This method requires the DEGG signals only during optimization.In operation, the proposed pitch detector marks glottal closures based strictly on the acoustical speech w a v eform.In addition to the algorithm development, performance comparison results are presented.

Extended permutation (EP) filters are defined and analyzed. In particular, we focus on extended permutation rank selection (EPRS) filters. These filters are constrained to output an order statistic from an extended observation vector. This extended vector includes N observation samples and K statistics that are functions of the observation samples. The rank permutations from selected samples in this extended observation vector are used as the basis for selecting an order statistic output. We show that by including the sample mean in the extended observation vector, the filters exhibit excellent edge enhancement properties. We also show that several previously defined classes of rank-order-based edge enhancers (CS, LUM, and WMMR sharpeners) can be formulated as subclasses of EPRS filters. These sharpening subclasses are in addition to the smoothing subclasses, which include rank conditioned rank selection, permutation stack, and weighted order statistic filters. Thus, this novel class of filters provides a broad framework within which many rank-order-based smoothers and edge enhancers can be unified. Edge enhancement properties are developed and an L(n) norm EPRS filter optimization procedure is presented. Finally, extensive computer simulation results are presented, comparing the performance of EPRS and other sharpening filters in edge enhancement applications.

Extended permutation (EP) filters are defined and analyzed in this paper. In particular, we focus on extended permutation rank selection (EPRS) filters. These filters are constrained to output an order statistic from an extended observation vector. This extended vector includes N observation samples and K statistics that are functions of the observation samples. By selecting an appropriate extended observation space, we show that the EPRS filters can be designed to have excellent edge enhancement characteristics. Moreover, the EPRS filters can perform edge enhancement in the presence of noise making them a powerful filter class.

We have introduced and analyzed a new class of adaptive nonlinear filters referred to as partition-based linear (Pl) filters. The operation of those filters depends on partitioning the observation space in some fashion. Specifically, we have used here scaler quantization as an example to illustrate the concept of partitioning the observation space. Each partition is then assigned an output based on linear combinations of observed samples in a moving window of finite length N. The filters are shown to exhibit appealing robustness. Simulations include a novel approach to estimating response-to-response variations in evoked potentials (EP), buried in the on-going electroencephalogram (EEG). Unlike the multi-channel filters currently used in EP estimation, the Pl filters do not require a separate electrode to provide a reference signal. In addition, no repetition of the stimulus is needed and the time of the stimulus need not be known.

Extended permutation (EP) filters are defined and analyzed in this paper. In particular, we focus on extended permutation rank selection (EPRS) filters. These filters are constrained to output an order statistic from an extended observation vector. This extended vector includes N observation samples and K statistics that are functions of the observation samples. By selecting an appropriate extended observation space, we show that the EPRS filters can be designed to have excellent edge enhancement characteristics. Moreover, the EPRS filters contain several previously defined edge enhancing filters as a subset and can perform edge enhancement in the presence of noise.

We introduce and analyze a new class of nonlinear filters called permutation weighted order statistic (PWOS) filters. These filters extend the concept of weighted order statistic (WOS) filters, in which filter weights associated with the input samples are used to replicate the corresponding samples, and an order statistic is chosen as the filter output. PWOS filters replicate each input sample according to weights determined by the temporal-order and rank-order of samples within a window. Hence, PWOS filters are in essence time-varying WOS filters. By varying the amount of temporal-rank order information used in selecting the output for a given observation window size, we obtain a wide range of filters that are shown to comprise a complete lattice structure. At the simplest level in the lattice, PWOS filters reduce to the well-known WOS filter, but for higher levels in the lattice, the obtained selection filters can model complex nonlinear systems and signal distortions. It is shown that PWOS filters are realizable by a N! piecewise linear threshold logic gate where the coefficients within each partition can be easily optimized using stack filter theory. Simulations are included to show the advantages of PWOS filters for the processing of image and video signals.

A class of nonlinear filters called rank conditioned rank selection (RCRS) filters is developed and analyzed in this paper. The RCRS filters are developed within the general framework of rank selection (RS) filters, which are filters constrained to output an order statistic from the observation set. Many previously proposed rank order based filters can be formulated as RS filters. The only difference between such filters is in the information used in deciding which order statistic to output. The information used by RCRS filters is the ranks of selected input samples, hence the name rank conditioned rank selection filters. The number of input sample ranks used is referred to as the order of the RCRS filter. The order can range from zero to the number of samples in the observation window, giving the filters valuable flexibility. Low-order filters can give good performance and are relatively simple to optimize and implement. If improved performance is demanded, the order can be increased but at the expense of filter simplicity. In this paper, many statistical and deterministic properties of the RCRS filters are presented. A procedure for optimizing over the class of RCRS filters is also presented. Finally, extensive computer simulation results that illustrate the performance of RCRS filters in comparison with other techniques in image restoration applications are presented.

The author introduces the combination stack filter, or C-stack filter, as a generalization of stack filters. This class of filters was developed in order to incorporate multilevel information into threshold level Boolean-function based filters (stack filters). C-stack filters generalize stack filters by performing a one-to-one transformation on the input sequence prior to Boolean filtering. The transformed sequence contains both the rank and spatial information not lost in the thresholding process. The knowledge of the rank information at the threshold level increases the number of filters possible and their performance. A brief review of stack and generalized stack filters is presented as well as several simulation results comparing stack, generalized stack, and C-stack filters.>