In this paper, we propose a new benchmark called "Archaeological Piece Grouping." In the field of archaeology, it is common for broken archaeological pieces, such as artifact fragments, to be mixed. Archaeologists often spend significant time distinguishing these pieces and categorizing them into different groups. Our benchmark introduces a novel, comprehensive dataset named ArcPie, along with new evaluation metrics for this task. Additionally, we propose a new framework called "3D Probabilistic Graph Search" (3DPGS) to address the problem of grouping mixed archaeological pieces. This framework includes a relation network designed to learn the relationships among all the input 3D pieces. Utilizing the relationships learned, our framework generates a probabilistic matching graph that describes the affinity of any two pieces. We also introduce a novel search algorithm to identify groups according to this matrix. Our framework significantly outperforms other baselines.

We introduce Cluster Contrast (CueCo), a novel approach to unsupervised visual representation learning that effectively combines the strengths of contrastive learning and clustering methods. Inspired by recent advancements, CueCo is designed to simultaneously scatter and align feature representations within the feature space. This method utilizes two neural networks, a query and a key, where the key network is updated through a slow-moving average of the query outputs. CueCo employs a contrastive loss to push dissimilar features apart, enhancing inter-class separation, and a clustering objective to pull together features of the same cluster, promoting intra-class compactness. Our method achieves 91.40% top-1 classification accuracy on CIFAR-10, 68.56% on CIFAR-100, and 78.65% on ImageNet-100 using linear evaluation with a ResNet-18 backbone. By integrating contrastive learning with clustering, CueCo sets a new direction for advancing unsupervised visual representation learning.

Few-shot learning investigates how to solve novel tasks given limited labeled data. Exploiting unlabeled data along with the limited labeled has shown substantial improvement in performance. In this work we propose a novel algorithm that exploits unlabeled data in order to improve the performance of few-shot learning. We focus on transductive few-shot inference, where the entire test set is available at inference time, and semi-supervised few-shot learning where unlabeled data are available and can be exploited. Our algorithm starts by leveraging the manifold structure of the labeled and unlabeled data in order to assign accurate pseudo-labels to the unlabeled data. Iteratively, it selects the most confident pseudo-labels and treats them as labeled improving the quality of pseudo-labels at every iteration. Our method surpasses or matches the state of the art results on four benchmark datasets, namely mini ImageNet, tiered ImageNet, CUB and CIFAR-FS, while being robust over feature pre-processing and the quantity of available unlabeled data. Furthermore, we investigate the setting where the unlabeled data contains data from distractor classes and propose ideas to adapt our algorithm achieving new state of the art performance in the process. Specifically, we utilize the unnormalized manifold class similarities obtained from label propagation for pseudo-label cleaning and exploit the uneven pseudo-label distribution between classes to remove noisy data. The publicly available source code can be found at https://github.com/MichalisLazarou/iLPC . • Unlabeled examples can improve few-shot learning performance significantly. • A novel algorithm for transductive and semi-supervised few-shot learning is proposed. • Accurate predictions for unlabeled examples are obtained by label propagation. • Label cleaning is used to select the most confident pseudo-labeled examples. • State of the art performance achieved in multiple few-shot learning settings.

This paper proposes a novel task named “3D part grouping”. Suppose there is a mixed set containing scattered parts from various shapes. This task requires algorithms to find out every possible combination among all the parts. To address this challenge, we propose the so called Gradient Field-based Auto-Regressive Sampling framework (G-FARS) tailored specifically for the 3D part grouping task. In our framework, we design a gradient-field-based selection graph neural network (GNN) to learn the gradients of a log conditional probability density in terms of part selection, where the condition is the given mixed part set. This innovative approach, implemented through the gradient-field-based selection GNN, effectively captures complex relationships among all the parts in the input. Upon completion of the training process, our framework becomes capable of autonomously grouping 3D parts by iteratively selecting them from the mixed part set, leveraging the knowledge acquired by the trained gradient-field-based selection GNN. Our code is available at: https://github.com/J-F-Cheng/G-FARS-3DPartGrouping.

Transductive few-shot learning algorithms have showed substantially superior performance over their inductive counterparts by leveraging the unlabeled queries at inference. However, the vast majority of transductive methods are evaluated on perfectly class-balanced benchmarks. It has been shown that they undergo remarkable drop in performance under a more realistic, imbalanced setting.To this end, we propose a novel algorithm to address imbalanced transductive few-shot learning, named Adaptive Manifold. Our algorithm exploits the underlying manifold of the labeled examples and unlabeled queries by using manifold similarity to predict the class probability distribution of every query. It is parameterized by one centroid per class and a set of manifold parameters that determine the manifold. All parameters are optimized by minimizing a loss function that can be tuned towards class-balanced or imbalanced distributions. The manifold similarity shows substantial improvement over Euclidean distance, especially in the 1-shot setting.Our algorithm outperforms all other state of the art methods in three benchmark datasets, namely miniImageNet, tieredImageNet and CUB, and two different backbones, namely ResNet-18 and WideResNet-28-10. In certain cases, our algorithm outperforms the previous state of the art by as much as 4.2%. The publicly available source code can be found in https://github.com/MichalisLazarou/AM

Few-shot learning addresses the issue of classifying images using limited labeled data. Exploiting unlabeled data through the use of transductive inference methods such as label propagation has been shown to improve the performance of few-shot learning significantly. Label propagation infers pseudo-labels for unlabeled data by utilizing a constructed graph that exploits the underlying manifold structure of the data. However, a limitation of the existing label propagation approaches is that the positions of all data points are fixed and might be sub-optimal so that the algorithm is not as effective as possible. In this work, we propose a novel algorithm that adapts the feature embeddings of the labeled data by minimizing a differentiable loss function optimizing their positions in the manifold in the process. Our novel algorithm, Adaptive Anchor Label Propagation, outperforms the standard label propagation algorithm by as much as 7% and 2% in the 1-shot and 5-shot settings respectively. We provide experimental results highlighting the merits of our algorithm on four widely used few-shot benchmark datasets, namely miniImageNet, tieredImageNet, CUB and CIFAR-FS and two commonly used backbones, ResNet12 and WideResNet-28-10. The source code can be found at https://github.com/MichalisLazarou/A2LP.

Co-salient object detection (CoSOD) together with the rapid development of deep learning has led to substantial progress in recent years. However, the feature aggregation between group feature representation and individual feature representation is still a challenging issue. In this work, we propose a novel adaptive intra-group aggregation (AIGA) method, which provides a new perspective to investigate the interaction relationship between group and single-image features and aggregate these features in an adaptive way. A novel scale-aware loss is proposed to help the model capture the scale prior of different groups and discriminatively process groups during the training phase. Extensive experiments demonstrate that the proposed method can effectively improve the performance without increasing extra parameters and achieve better accuracy on three prevalent benchmarks.

Pedestrian detection is a key task in intelligent video surveillance systems which requires both fast inference and high detection accuracy. Although single-stage deep learning pedestrian detectors have achieved relatively high detection accuracy with simpler architecture and less inference time, their performance is limited compared to two-stage methods. The reason is the lack of scale-aware features without the assistance of proposal regions. To overcome this, a multi-scale deformable transformer encoder-based module is proposed. It can extract the sparse important features at deformable sampling locations from multiple levels. The proposed architecture significantly improves the performance compared to the baseline center and scale prediction method on both Caltech and Citypersons datasets. It even outperforms the state-of-the-art two-stage methods in detecting heavily occluded pedestrians on Citypersons validation set.

Vision transformers have demonstrated remarkable performance on a variety of computer vision tasks. In this paper, we illustrate the effectiveness of the deformable vision transformer for single-stage pedestrian detection and propose a spatial and multi-scale feature enhancement module, which aims to achieve the optimal balance between speed and accuracy. Performance improvement with vision transformers on various commonly used single-stage structures is demonstrated. The design of the proposed architecture is investigated in depth. Comprehensive comparisons with state-of-the-art single- and two-stage detectors on different pedestrian datasets are performed. The proposed detector achieves leading performance on Caltech and Citypersons datasets among single- and two-stage methods using fewer parameters than the baseline. The log-average miss rates for Reasonable and Heavy are decreased to 2.6% and 28.0% on the Caltech test set, and 10.9% and 38.6% on the Citypersons validation set, respectively. The proposed method outperforms SOTA two-stage detectors in the Heavy subset on the Citypersons validation set with considerably faster inference speed.

Few-shot learning addresses the challenge of learning how to address novel tasks given not just limited supervision but limited data as well. An attractive solution is synthetic data generation. However, most such methods are overly sophisticated, focusing on high-quality, realistic data in the input space. It is unclear whether adapting them to the few-shot regime and using them for the downstream task of classification is the right approach. Previous works on synthetic data generation for few-shot classification focus on exploiting complex models, e.g. a Wasserstein GAN with multiple regularizers or a network that transfers latent diversities from known to novel classes.We follow a different approach and investigate how a simple and straightforward synthetic data generation method can be used effectively. We make two contributions, namely we show that: (1) using a simple loss function is more than enough for training a feature generator in the few-shot setting; and (2) learning to generate tensor features instead of vector features is superior. Extensive experiments on miniImagenet, CUB and CIFAR-FS datasets show that our method sets a new state of the art, outperforming more sophisticated few-shot data augmentation methods. The source code can be found at https://github.com/MichalisLazarou/TFH_fewshot.

Salient object detection (SOD) aims at locating the most significant object within a given image. In recent years, great progress has been made in applying SOD on many vision tasks. The depth map could provide additional spatial prior and boundary cues to boost the performance. Combining the depth information with image data obtained from standard visual cameras has been widely used in recent SOD works, however, introducing depth information in a suboptimal fusion strategy may have negative influence in the performance of SOD. In this paper, we discuss about the advantages of the so-called progressive multi-scale fusion method and propose a mask-guided feature aggregation module (MGFA). The proposed framework can effectively combine the two features of different modalities and, furthermore, alleviate the impact of erroneous depth features, which are inevitably caused by the variation of depth quality. We further introduce a mask-guided refinement module (MGRM) to complement the high-level semantic features and reduce the irrelevant features from multi-scale fusion, leading to an overall refinement of detection. Experiments on five challenging benchmarks demonstrate that the proposed method outperforms 11 state-of-the-art methods under different evaluation metrics.

Few-shot learning amounts to learning representations and acquiring knowledge such that novel tasks may be solved with both supervision and data being limited. Improved performance is possible by transductive inference, where the entire test set is available concurrently, and semi-supervised learning, where more unlabeled data is available.Focusing on these two settings, we introduce a new algorithm that leverages the manifold structure of the labeled and unlabeled data distribution to predict pseudo-labels, while balancing over classes and using the loss value distribution of a limited-capacity classifier to select the cleanest labels, iteratively improving the quality of pseudo-labels. Our solution surpasses or matches the state of the art results on four benchmark datasets, namely miniImageNet, tieredImageNet, CUB and CIFAR-FS, while being robust over feature space pre-processing and the quantity of available data. The publicly available source code can be found in https://github.com/MichalisLazarou/iLPC

Pedestrian detection methods have been significantly improved with the development of deep convolutional neural networks. Nevertheless, detecting ismall-scaled pedestrians and occluded pedestrians remains a challenging problem. In this paper, we propose a pedestrian detection method with a couple-network to simultaneously address these two issues. One of the sub-networks, the gated multi-layer feature extraction sub-network, aims to adaptively generate discriminative features for pedestrian candidates in order to robustly detect pedestrians with large variations on scale. The second sub-network targets on handling the occlusion problem of pedestrian detection by using deformable regional region of interest (RoI)-pooling. We investigate two different gate units for the gated sub-network, namely, the channel-wise gate unit and the spatio-wise gate unit, which can enhance the representation ability of the regional convolutional features among the channel dimensions or across the spatial domain, repetitively. Ablation studies have validated the effectiveness of both the proposed gated multi-layer feature extraction sub-network and the deformable occlusion handling sub-network. With the coupled framework, our proposed pedestrian detector achieves promising results on both two pedestrian datasets, especially on detecting small or occluded pedestrians. On the CityPersons dataset, the proposed detector achieves the lowest missing rates (i.e. 40.78% and 34.60%) on detecting small and occluded pedestrians, surpassing the second best comparison method by 6.0% and 5.87%, respectively.

Pedestrian detection methods have been significantly improved with the development of deep convolutional neural networks. Nevertheless, it remains a challenging problem how to robustly detect pedestrians of varied sizes and with occlusions. In this paper, we propose a gated multi-layer convolutional feature extraction method which can adaptively generate discriminative features for candidate pedestrian regions. The proposed gated feature extraction framework consists of squeeze units, gate units and concatenation layers which perform feature dimension squeezing, feature manipulation and features combination from multiple CNN layers, respectively. We proposed two different gate models that can manipulate the regional feature maps in a channel-wise selection manner and a spatial-wise selection manner, respectively. Experiments on the challenging CityPersons dataset demonstrate the effectiveness of the proposed method, especially on detecting small-size and occluded pedestrians.

Convolutional neural networks (CNNs) have resurged lately due to their state-of-the-art performance in various disciplines, such as computer vision , audio and text processing. However, CNNs have not been widely employed for remote sensing applications. In this paper, we propose a CNN architecture , named Modular-CNN, to improve the performance of building detectors that employ Histogram of Oriented Gradients (HOG) and Local Binary Patterns (LBP) in a remote sensing dataset. Additionally, we propose two improvements to increase the classification accuracy of Modular-CNN. The first improvement combines the power of raw and normalised features, while the second one concerns the Euler transformation of feature vectors. We demonstrate the effectiveness of our proposed Modular-CNN and the novel improvements in remote sensing and other datasets in a comparative study with other state-of-the-art methods.

Invasive insect pests and fungi, which are introduced accidentally to forests and affect tree growth and survival, constitute a serious threat for the forests and trees acting on climate change and its impacts. Thus, the need for early and accurate health determination process of forest regions, has significantly increased the interest in automatic monitoring methods. In this paper, in order to overcome the great variety of trees’ characteristics and forests’ heterogeneity that affects the diversity of their color and texture making the detection of diseases a difficult task, a methodology for individual tree detection applying energy minimization, visualizing HOG features across tree canopies and using a dynamically clustering method is proposed. Then, in order to achieve classification based on their health condition, multi-pyramid textural features are proposed and extracted. The experimental results presented use images of a forest area in Greece that include fir trees and show the great potential of the proposed methodology.

Given the urgent priority around protecting the forests and limiting the impacts of the climate change, the constant monitoring of forests towards the achievement of accurate and timely detection of infestations and the catastrophic action of invasive insects, pests and fungi is an important and challenging task. More precisely, new species of insects that are introduced or already existing insect species whose population multiply uncontrollably into the forest area, affect tree growth, their survival, as well as the quality of forest biomass and constitute a serious threat to the mechanisms of such forest ecosystems. Thus, new concepts are needed that will overcome difficulties faced by existing remote sensing techniques and that would allow the timely and accurate health determination process of forest regions, assisting scientists and authorities to take action in order to protect the forests. In this paper, we propose a monitoring approach, which uses high resolution RGB aerial images and combines different Region Convolution Neural Networks (R-CNNs) architectures, namely Faster R-CNN and Mask R-CNN and fuses their bounding box outcomes in order to more accurately localize candidate infected trees’ regions whilst increasing the number of the candidate trees that have been detected as infected. Subsequently, the candidate detected trees are modelled through the higher order linear dynamical systems (h-LDS) and descriptors are extracted for each candidate region. Finally, the h-LDS descriptors are classified using an SVM classifier for the estimation of the infected trees. The study area includes parts of the suburban pine forest of Thessaloniki city (Greece) named Seich Sou, which suffers the last months an infestation of high significance and intensity by a bark and wood destroying insect (Tomicus piniperda). Although this insect was recorded in the specific ecosystem many years ago, its population increased uncontrollably after the degradation of the ecosystem due to human intervention and lack of protection and management strategy. Experimental results, through their outperforming existing state-of-the-art algorithms, demonstrate high potential and perspectives of the proposed methodology of low cost and time consumed, to contribute to the sustainable management, protection and recovery of a forest ecosystem.

Hyper-spectral satellite imagery, consisting of multiple visible or infrared bands, is extremely dense and weighty for deep operations. Regarding problems related to vegetation as, more specifically, tree segmentation, it is difficult to train deep architectures due to lack of large-scale satellite imagery. In this paper, we compare the success of different single channel indices, which are constructed from multiple bands, for the purpose of tree segmentation in a deep convolutional neural network (CNN) architecture. The utilized indices are either hand-crafted such as excess green index (ExG) and normalized difference vegetation index (NDVI) or reconstructed from the visible bands using feature space transformation methods such as principle component analysis (PCA). For comparison, these features are fed to an identical CNN architecture, which is a standard U-Net-based symmetric encoder-decoder design with hierarchical skip connections and the segmentation success for each single index is recorded. Experimental results show that single bands, which are constructed from the vegetation indices and space transformations, can achieve similar segmentation performances as compared to that of the original multi-channel case.

Human and object detection is a continuously explored subject in tracking and navigation applications as well as within the machine vision community. More precisely, in navigation applications that are designed for robotics purposes or in order to support car drivers, the real-time detection of presence of humans and other objects is an important and challenging task. Specifically, human and object detection is a significant part of a human-computer collaboration in the sense of allowing automatic navigation and control systems to obtain a better model of the real world. Thus, new concepts are needed that will overcome difficulties faced by existing human detection approaches and will allow navigation systems to perform accurately. This paper offers a general classification scheme combining different detector systems towards navigation tasks. To evaluate the performance of the proposed methodology we used INRIA and PASCAL VOC 2007 datasets. Experimental results show that the combination of different image feature descriptors, classifier models and deep learning techniques are advantageous for human detection.

Phase correlation (PC) is widely employed by several sub-pixel motion estimation techniques in an attempt to accurately and robustly detect the displacement between two images. To achieve sub-pixel accuracy, these techniques employ interpolation methods and function-fitting approaches on the cross-correlation function derived from the PC core. However, such motion estimation techniques still present a lower bound of accuracy that cannot be overcome. To allow room for further improvements, we propose in this paper the enhancement of the sub-pixel accuracy of motion estimation techniques by employing a completely different approach: the concept of motion magnification. To this end, we propose the novel phase amplified correlation (PAC) that integrates motion magnification between two compared images inside the phase correlation part of frequencybased motion estimation algorithms and thus directly substitutes the PC core. The experimentation on magnetic resonance (MR) images and real video sequences demonstrates the ability of the proposed PAC core to make subtle motions highly distinguishable and improve the sub-pixel accuracy of frequency-based motion estimation techniques.

Detection of small targets, more specifically cars, in aerial images of urban scenes, has various applications in several domains, such as surveillance, military, remote sensing, and others. This is a tremendously challenging problem, mainly because of the significant interclass similarity among objects in urban environments, e.g., cars and certain types of nontarget objects, such as buildings' roofs and windows. These nontarget objects often possess very similar visual appearance to that of cars making it hard to separate the car and the noncar classes. Accordingly, most past works experienced low precision rates at high recall rates. In this paper, a novel framework is introduced that achieves a higher precision rate at a given recall than the state of the art. The proposed framework adopts a sliding-window approach and it consists of four stages, namely, window evaluation, extraction and encoding of features, classification, and postprocessing. This paper introduces a new way to derive descriptors that encode the local distributions of gradients, colors, and texture. Image descriptors characterize the aforementioned cues using adaptive cell distributions, wherein the distribution of cells within a detection window is a function of its dominant orientation, and hence, neither the rotation of the patch under examination nor the computation of descriptors at different orientations is required. The performance of the proposed framework has been evaluated on the challenging Vaihingen and Overhead Imagery Research data sets. Results demonstrate the superiority of the proposed framework to the state of the art.

In this work, the 3D flow estimation problem is formulated in the 4D spatiotemporal frequency domain, and it is shown that 3D motion manifests itself as energy concentration along hyper-planes in that domain. Based on this, the construction and use of appropriate directional multidimensional “steerable” filters, which can extract directional energy in spacetime, is proposed. Steerable filters have been constructed for up to 3 dimensions. We extend the relevant mathematical definitions to multiple dimensions and formulate filter-based algorithms for 3D flow estimation. Experimental results on simulated and real data verify the efficiency of the algorithms.

Robust and efficient detection of cars in urban scenes has many useful applications. This paper introduces a framework for car detection from high-resolution satellite images, wherein a novel extended image descriptor is used to depict the geometric, spectral and colour distribution properties of cars. The proposed framework is based on a sliding-window detection approach and it begins with a pre-prepossessing stage, which discards detection windows that are very unlikely to contain cars, e.g., plain areas and vegetation, followed by the computation of a concatenated feature vector of Histogram of Oriented Gradients, Fourier and truncated Pyramid Colour Self-Similarity image descriptors that is then fed to a pre-trained linear Support Vector Machine classifier to discriminate between the feature and non-feature subspaces. For post-processing, a non-maximum suppression technique is used to eliminate multiple detections. The performance of the proposed framework has been assessed on the Vaihingen dataset and results show that it exceeds the performance of the current state-of-the-art car detection algorithms.

The derivation of habitat maps is enhanced if land cover maps are used as basis for the mapping procedure. In this study, a supervised learning framework is proposed to perform object-based classification to General Habitat Categories. A Land Cover Classification System map is used as basis, and an approach to generate numerical features from the object land cover class names and attributes is introduced. An additional number of spectral, morphological, and topological features are extracted from very high resolution satellite imagery and classification accuracies up to 80.4% for 14 classes are reached. Inclusion of LiDAR (Light Detection And Ranging) data or proposed texture analysis features, improve accuracies to 86% and around 83%, respectively, with the latter proving as promising surrogates of LiDAR data features. The method outperformed rule-based approaches, indicating its potential in accurate and labor- and time-efficient habitat classification.

In this paper, a new feature selection method is used, in combination with a semantic model vector video representation, in order to enumerate the key semantic evidences of an event in a video signal. In particular, a set of semantic concept detectors is firstly used for estimating a model vector for each video signal, where each element of the model vector denotes the degree of confidence that the respective concept is depicted in the video. Then, a novel feature selection method is learned for each event of interest. This method is based on exploiting the first two eigenvectors derived using the eigenvalue formulation of the mixture subclass discriminant analysis. Subsequently, given a video-event pair, the proposed method jointly evaluates the significance of each concept for the detection of the given event and the degree of confidence with which this concept is detected in the given video, in order to decide which concepts provide the strongest evidence in support of the provided video-event link. Experimental results using a video collection of TRECVID demonstrate the effectiveness of the proposed video event recounting method.

In this paper, complex video events are learned and detected using a novel subclass recoding error-correcting outputs (SRECOC) design. In particular, a set of pre-trained concept detectors along different low-level visual feature types are used to provide a model vector representation of video signals. Subsequently, a subclass partitioning algorithm is used to divide only the target event class to several subclasses and learn one subclass detector for each event subclass. The pool of the subclass detectors is then combined under a SRECOC framework to provide a single event detector. This is achieved by first exploiting the properties of the linear loss-weighted decoding measure in order to derive a probability estimate along the different event subclass detectors, and then utilizing the sum probability rule along event subclasses to retrieve a single degree of confidence for the presence of the target event in a particular test video. Experimental results on the large-scale video collections of the TRECVID Multimedia Event Detection (MED) task verify the effectiveness of the proposed method. Moreover, the effect of weak or strong concept detectors on the accuracy of the resulting event detectors is examined.

In this paper, a theoretical link between mixture subclass discriminant analysis (MSDA) and a restricted Gaussian model is first presented. Then, two further discriminant analysis (DA) methods, i.e., fractional step MSDA (FSMSDA) and kernel MSDA (KMSDA) are proposed. Linking MSDA to an appropriate Gaussian model allows the derivation of a new DA method under the expectation maximization (EM) framework (EM-MSDA), which simultaneously derives the discriminant subspace and the maximum likelihood estimates. The two other proposed methods generalize MSDA in order to solve problems inherited from conventional DA. FSMSDA solves the subclass separation problem, that is, the situation in which the dimensionality of the discriminant subspace is strictly smaller than the rank of the inter-between-subclass scatter matrix. This is done by an appropriate weighting scheme and the utilization of an iterative algorithm for preserving useful discriminant directions. On the other hand, KMSDA uses the kernel trick to separate data with nonlinearly separable subclass structure. Extensive experimentation shows that the proposed methods outperform conventional MSDA and other linear discriminant analysis variants.

This study describes a method for tracking objects through scale and occlusion. The technique presented is based on the mean shift algorithm, which provides an efficient way to track objects based on their colour characteristics. A novel and efficient method is derived for tracking through changes in the target scale, where an object of interest moves away or towards the camera and therefore appears to change size in the image plane. The method works by interleaving spatial mean shift iterations with scale iterations. It is shown that this method is considerably more efficient than other methods and possesses other advantages too. It is also demonstrated that the Bhattacharyya coefficient, a histogram similarity metric that is used in the mean shift framework, can be used to reliably detect when target occlusion occurs. In such situations, the motion of an object can be extrapolated to give an accurate estimate of its position. This is used as the basis of a technique for tracking through occlusion. Experimental results are presented on data from various scenarios.

In this letter, linear subclass support vector machines (LSSVMs) are proposed that can efficiently learn a piecewise linear decision function for binary classification problems. This is achieved using a nongaussianity criterion to derive the subclass structure of the data, and a new formulation of the optimization problem that exploits the subclass information. LSSVMs provide low computation cost during training and evaluation, and offer competitive recognition performance in comparison to other popular SVM-based algorithms. Experimental results on various datasets confirm the advantages of LSSVMs.

We propose a robust approach to discriminant kernel-based feature extraction for face recognition and verification. We show, for the first time, how to perform the eigen analysis of the within-class scatter matrix directly in the feature space. This eigen analysis provides the eigenspectrum of its range space and the corresponding eigenvectors as well as the eigenvectors spanning its null space. Based on our analysis, we propose a kernel discriminant analysis (KDA) which combines eigenspectrum regularization with a feature-level scheme (ER-KDA). Finally, we combine the proposed ER-KDA with a nonlinear robust kernel particularly suitable for face recognition/verification applications which require robustness against outliers caused by occlusions and illumination changes. We applied the proposed framework to several popular databases (Yale, AR, XM2VTS) and achieved state-of-the-art performance for most of our experiments.

The aim of this paper is to provide an algorithm for image fusion which combines the techniques of Chebyshev polynomial (CP) approximation and independent component analysis (ICA), based on the regional information of input images. We present a region-based method that combines the merits of both techniques. It utilises segmentation to identify edges, texture and other important features in the input image and subsequently apply the different fusion methods according to regions. The proposed method exhibits better perceptual performance than individual CP and ICA fusion approaches especially in noise corrupted images.

We address the problem of subpixel registration of images assumed to be related by a pure translation. We present a method which extends gradient correlation to achieve subpixel accuracy. Our scheme is based on modeling the dominant singular vectors of the 2-D gradient correlation matrix with a generic kernel which we derive by studying the structure of gradient correlation assuming natural image statistics. Our kernel has a parametric form which offers flexibility in modeling the functions obtained from various types of image data. We estimate the kernel parameters, including the unknown subpixel shifts, using the Levenberg-Marquardt algorithm. Experiments with LANDSAT and MRI data show that our scheme outperforms recently proposed state-of-the-art phase correlation methods.

This report documents in detail the research carried out by the author throughout his first year. The paper presents a novel method for fusing images in a domain concerning multiple sensors and modalities. Using Chebyshev polynomials as basis functions, the image is decomposed to perform fusion at feature level. Results show favourable performance compared to previous efforts on image fusion, namely ICA and DT-CWT, in noise affected images. The work presented here aims at providing a novel framework for future studies in image analysis and may introduce innovations in the fields of surveillance, medical imaging and remote sensing.

We present a robust FFT-based approach to scale-invariant image registration. Our method relies on FFT-based correlation twice: once in the log-polar Fourier domain to estimate the scaling and rotation and once in the spatial domain to recover the residual translation. Previous methods based on the same principles are not robust. To equip our scheme with robustness and accuracy, we introduce modifications which tailor the method to the nature of images. First, we derive efficient log-polar Fourier representations by replacing image functions with complex gray-level edge maps. We show that this representation both captures the structure of salient image features and circumvents problems related to the low-pass nature of images, interpolation errors, border effects, and aliasing. Second, to recover the unknown parameters, we introduce the normalized gradient correlation. We show that, using image gradients to perform correlation, the errors induced by outliers are mapped to a uniform distribution for which our normalized gradient correlation features robust performance. Exhaustive experimentation with real images showed that, unlike any other Fourier-based correlation techniques, the proposed method was able to estimate translations, arbitrary rotations, and scale factors up to 6.

A fast and robust gradient-based motion estimation technique which operates in the frequency domain is presented. The algorithm combines the natural advantages of a good feature selection offered by gradient-based methods with the robustness and speed provided by FFT-based correlation schemes. Experimentation with real images taken from a popular database showed that, unlike any other Fourier-based techniques, the method was able to estimate translations, arbitrary rotations and scale factors in the range 4-6.

In this paper, the problem of moment-based shape orientation and symmetry classification is jointly considered. A generalization and modification of current state-of-the-art geometric moment-based functions is introduced. The properties of these functions are investigated thoroughly using Fourier series analysis and several observations and closed-form solutions are derived. We demonstrate the connection between the results presented in this work and symmetry detection principles suggested from previous complex moment-based formulations. The proposed analysis offers a unifying framework for shape orientation/symmetry detection. In the context of symmetry classification and matching, the second part of this work presents a frequency domain method, aiming at computing a robust moment-based feature set based on a true polar Fourier representation of image complex gradients and a novel periodicity detection scheme using subspace analysis. The proposed approach removes the requirement for accurate shape centroid estimation, which is the main limitation of moment-based methods, operating in the image spatial domain. The proposed framework demonstrated improved performance, compared to state-of-the-art methods.

A novel frequency domain approach to roto-translation estimation is presented. The baseline gradient cross-correlation method is extended to handle rotations. A key feature of the proposed scheme is the ability to achieve good performance in the presence of both large translations and rotations as well as noise, a scenario for which other Fourier-based methods typically fail. Robustness and accuracy in conjunction with computational efficiency, offered by the frequency domain formulation, make the algorithm useful in a number of image processing tasks such as image registration. 1

A frequency domain approach for the detection of symmetries in real images is presented. Our framework is based on recent state-of-the-art research where motion estimation techniques are employed to sequentially determine all the associated parameters. In particular, we introduce several modifications regarding the order of symmetry estimation and the detection of the axes of possible bilateral symmetry. Preliminary results demonstrate the efficiency of our approach.

The problem of shape-based recognition of objects under affine transformations is considered. We focus on the construction of a robust and highly discriminative affine invariant function that can be used for within-class object recognition applications. Using the boundaries of the objects of interest, a training scheme, based on principal component analysis (PCA), is proposed to derive a set of basis functions with desired properties. The derived bases are then used for the construction of a novel affine invariant function. The proposed invariant function is evaluated for the problem of aircraft silhouette identification and appears to achieve comparable performance to a popular wavelet-based affine invariant function. At the same time, the proposed framework is much simpler than that based on wavelet analysis.

The purpose of the Applied Multi-dimensional Fusion Project is to investigate the benefits that data fusion and related techniques may bring to future military Intelligence Surveillance Target Acquisition and Reconnaissance systems. In the course of this work, it is intended to show the practical application of some of the best multi-dimensional fusion research in the UK. This paper highlights the work done in the area of multi-spectral synthetic data generation, super-resolution, joint fusion and blind image restoration, multi-resolution target detection and identification and assessment measures for fusion. The paper also delves into the future aspirations of the work to look further at the use of hyper-spectral data and hyper-spectral fusion. The paper presents a wide work base in multi-dimensional fusion that is brought together through the use of common synthetic data, posing real-life problems faced in the theatre of war. Work done to date has produced practical pertinent research products with direct applicability to the problems posed.

Image fusion systems aim at transferring ‘interesting’ information from the input sensor images to the fused image. The common assumption for most fusion approaches is the existence of a high-quality reference image signal for all image parts in all input sensor images. In the case that there are common degraded areas in at least one of the input images, the fusion algorithms cannot improve the information provided there, but simply convey a combination of this degraded information to the output. The authors propose a combined spatial-domain method of fusion and restoration in order to identify these common degraded areas in the fused image and use a regularized restoration approach to enhance the content in these areas. The proposed approach was tested on both multi-focus and multi-modal image sets and produced interesting results.

The problem of blind separation of statistically independent sources from instantaneous mixtures, using the efficient framework of independent component analysis (ICA), has been widely addressed in the literature. In this letter, the authors propose a sequential blind signal extraction algorithm that attempts to identify smooth sources in instantaneous mixtures. The approach incorporates smoothness constraints in the traditional negentropy cost function to extract smooth components, using an approximate second-order optimization method

A scheme, based on principal component analysis (PCA), is proposed that can be used for the recognition of 2-D planar shapes under affine transformations. A PCA step is first used to map the object boundary to its canonical form, reducing the problem of the nonuniform sampling of the object contour introduced by the affine transformation. Then, a PCA-based scheme is employed to train a set of basis functions on the signals extracted from the objects' boundaries. The derived bases are used to analyze the boundary locally. Based on the theory of invariants and local boundary analysis, a novel invariant function is constructed. The performance of the proposed framework is compared with a standard wavelet-based approach with promising results.

In this paper, we explore the problem of sound source separation and identification from a two-sensor instantaneous mixture. The estimation of the mixing and the sources is performed using Laplacian mixture models (LMM). The proposed algorithm fits the model using batch processing of the observed data and performs separation using either a hard or a soft decision scheme. An extension of the algorithm to online source separation, where the samples are arriving in a real-time fashion, is also presented. The online version demonstrates several promising source separation possibilities in the case of nonstationary mixing.

On the spatial domain, image fusion can be approached by estimating a set of fusion weights, that optimally measure the contribution of each pixel in source images to the fused one. Combining fusion weights with source images yields the fused result with improved visual perception. This paper aims to find these weights by minimising a constant-modulus (CM) cost function that describes the dispersion of the fused image. In order to accelerate convergence rate and avoid spurious solutions, we also introduce optimal learning rates while updating fusion weights. Experimental results reveal that our scheme provides comparable performance on fusing multifocus images to multi-scale wavelet methods, such as shift-invariant discrete wavelet transform (SI-DWT)

Image fusion can be viewed as a process that incorporates essential information from different modality sensors into a composite image. The use of bases trained using independent component analysis (ICA) for image fusion has been highlighted recently. Common fusion rules can be used in the ICA fusion framework with promising results. In this paper, the authors propose an adaptive fusion scheme, based on the ICA fusion framework, that maximises the sparsity of the fusion image in the transform domain

In this paper, we tackle the problem of image denoising. Given an image distorted by additive noise of unknown probability distribution, the intensity difference between the distorted and the original unknown images must be bounded. This idea is exploited here and we design a series of weighted adaptive local bounds based on local intensity information, such as the mean, variance, median and etc. The proposed method is tested and compared with other standard techniques, such as wavelet thresholding, for image denoising. Apart from the simplicity of implementation, the results are very encouraging, as far as both visual quality of the denoised images and quantitative metrics of improvement are concerned. More importantly it provides simultaneous denoising of mixed noise, which is not obtainable by using single conventional denoising methods.

The authors explore the use of Laplacian mixture models (LMMs) to address the overcomplete blind source separation problem in the case that the source signals are very sparse. A two-sensor setup was used to separate an instantaneous mixture of sources. A hard and a soft decision scheme were introduced to perform separation. The algorithm exhibits good performance as far as separation quality and convergence speed are concerned.

Physical signals are often non-stationary. We assume that images are non-stationary and each pixel is the output of a different set of two-dimensional autoregressive (AR) model coefficients; therefore, texture cannot be characterised by applying AR modelling techniques to the entire image. A multistage modelling technique, which appears in the literature as "endomorphic modelling", is applied to this kind of signal. The signal is divided into sub-sets of samples and each sub-set is modelled using an AR model. A new image is formed using the same AR model coefficient estimated from different blocks, then the new image is re-modelled using AR models again. This process ends when the variance of the estimated parameters is sufficiently small. The simulation results show that the variances obtained from the estimated coefficients at the second stage are much smaller than the ones from the first stage

A normalized robust mixed-norm (NRMN) algorithm for system identification in the presence of impulsive noise is introduced. The standard robust mixed-norm (RMN) algorithm exhibits slow convergence, requires a stationary operating environment, and employs a constant step-size that needs to be determined a priori. To overcome these limitations, the proposed NRMN algorithm introduces a time-varying learning rate and, thus, no longer requires a stationary environment, a major drawback of the RMN algorithm. The proposed NRMN exhibits increased convergence rate and substantially reduces the steady-state coefficient error, as compared to the least mean square (LMS), normalized LMS (NLMS), least absolute deviation (LAD), and RMN algorithm.

A DFT algorithm, the extended subband DFT (ESB-DFT), is presented, which overcomes the problem of low accuracy the subband DFT (SB-DFT) S.K. Mitra et al., (1990) O.V. Shentov et al., (1991) suffers from when applied in its approximate mode for signals with full spectrum. The proposed ESB-DFT is based upon the analysis of the input signal using a filter bank (FB) and the computation of its DFT from the lower length DFTs that correspond to the subband components resulting from the analysis. The proposed 1D scheme is extended to yield the 2D ESB DFT algorithm. The approximate mode of ESB-DFT is then introduced with a detailed error analysis, which shows that it is possible to reduce the computational complexity while obtaining a very accurate DFT estimate within a band of interest. The ESB-DFT is tested and compared to the SB-DFT for 1D and 2D signals. It is shown that ESB-DFT substantially improves the estimation accuracy of SB-DFT, while when applied in its full-mode yields the exact DFT of the input.

The convergence analysis of a maximally decimated (MD) adaptive filtering structure (AFS) employing a new adaptation scheme is presented. The structure was based on uniform perfect reconstruction (PR) filter banks (FB), their dual transmultiplexers (TMUX) and a model for the system under identification of interpolated FIR (IFIR) filters connected in parallel. A new algorithm was used for the adaptation of the subband filter coefficients employing a new power normalisation (PN) scheme introduced in Papoulis et al. (2002). This structure is briefly described here, the adaptive algorithm is presented and an extensive convergence analysis is then given. Reference then is made to the relevant work reported in Petraglia et al. (2000), namely to a structure that can be considered as the dual of the one presented here and employed an adaptive algorithm that is a simplified version of the one we proposed. The convergence analysis is then specialised for the adaptive scheme used in Petraglia. The rest of the paper is an extensive performance comparison of both structures and of the two adaptation schemes, through both theoretical and experimental results. These demonstrate the improvement in the performance our approach offers; highlight aspects of the theoretical analysis and further enlighten on the similarities and differences of the two approaches. A performance comparison with the full-band NLMS algorithm is given.

A maximally decimated (MD) structure is presented for adaptive filtering based on perfect reconstruction (PR) transmultiplexers (TMUXs) and a system representation that results from their dual filter banks (FBs). The structure is computationally very efficient adapting at the low data rate a number of coefficients that exceeds only by a small amount the length of the unknown system. For colored inputs the structure presents a higher convergence rate (CR) from the Full-band LMS algorithm. These characteristics are investigated through simulation experiments where the proposed structure is also compared with other recently proposed approaches and found to have a similar performance.

We propose a method for a minimum phase finite impulse response (FIR) equiripple filter design from a given linear phase FIR filter with the same amplitude response. We are concentrating on very high degree polynomials (transfer functions) for which factorisation procedures for root extraction are unreliable. The initial linear phase transfer functions are obtained by standard design algorithms and particularly in this study by the Remez algorithm. The approach taken involves the use of the Cauchy residue theorem applied to the logarithmic derivative of the appropriate curves. This leads into a set of parameters related directly to the polynomial coefficients, which facilitate the factorisation problem. The results of the proposed design scheme are very encouraging as far as robustness and computational complexity are concerned.

This paper forms a part of a series of previous studies we have undertaken, where the problem of nonlinear textured image modeling is examined. We assume that the observed "output" image is derived from a Volterra filter that is driven by a Gaussian input. Both the filter parameters and the input signal are unknown and therefore the problem can be classified as blind or unsupervised in nature. In the statistical approach to the solution of the above problem we seek for equations that relate the unknown parameters of the Volterra model with the statistical parameters of the "output" image to be modeled. These equations are highly nonlinear and their solution is achieved through a novel constrained optimisation formulation.

The objective of this paper is to produce a general formulation of an order reduction procedure for testing the stability of discrete time linear systems. The order reduction procedure involves a series of iterations and, at each step of the iteration process, the aim is to derive a new polynomial of order lower than the given one. The new polynomial serves as the input to the following iteration. A specific form of the formulation is considered in which first order auxiliary polynomials are employed in the order reduction process. There follows from this a new testing procedure which is computationally more efficient than the existing ones. Moreover the current methods appear as special cases of the new test. An extension is further proposed which employs second order auxiliary polynomials within the order reduction formulation. This second order form is, however, for all practical cases the limit to which such a procedure can be put.

A new method of incorporating local image characteristics into blind image restoration is proposed. The local variance of the degraded image is used as a measure of spatial activity, from which individual pixel bounds are determined. A parameter defined by the user controls the degree of smoothing. The local bounds define the solution more precisely than smoothness constraints on the image (including those that are spatially-adaptive), reducing the number of possible solutions and leading to a faster rate of convergence. Experimental results demonstrate the potential of this method as an alternative/supplement to smoothing constraints in blind image restoration.

In this contribution we propose a method for a minimum phase finite impulse response (FIR) filter design from a given linear phase FIR function with the same amplitude response. We concentrate on very high degree polynomials for which factorisation procedures for root extraction are unreliable. The approach taken involves using the Cauchy residue theorem applied to the logarithmic derivative of the transfer function. This leads to a set of parameters derivable directly from the polynomial coefficients which facilitate the factorisation problem. The concept is developed in a way that leads naturally to the celebrated Newton identities. In addition to solving the above problem, the results of the proposed design scheme are very encouraging as far as robustness and computational complexity are concerned.

In this paper, we present a regularized mixed norm multichannel image restoration algorithm. The problem of multichannel restoration using both within- and between-channel deterministic information is considered. For each channel a functional which combines the least mean squares (LMS), the least mean fourth (LMF), and a smoothing functional is proposed. We introduce a mixed norm parameter that controls the relative contribution between the LMS and the LMF, and a regularization parameter that defines the degree of smoothness of the solution, both updated at each iteration according to the noise characteristics of each channel. The novelty of the proposed algorithm is that no knowledge of the noise distribution for each channel is required, and the parameters mentioned above are adjusted based on the partially restored image.

In this paper, the problem of how to better estimate spatially adaptive intensity bounds for image restoration is addressed. When the intensity bounds are estimated from a degraded image, blurring leads to underestimation of the bounds in the edge and texture regions. Therefore, an iterative implementation of the restoration algorithm has been proposed in which the intensity bounds are re-estimated from the current image estimate. However, direct update of the bounds leads to over-smoothing in regions where the bounds are active. Furthermore, the resulting algorithm exhibits slow convergence. In this paper, alternative methods of initially estimating and updating the bounds are proposed, and the results for the fixed- and updated-bound implementations are compared. A method for estimation of the bound tightness parameter is also proposed.

A novel approach is taken for the estimation of the parameters of a Volterra model, which is based on constrained optimisation. The equations required for the determination of the Volterra kernels are formed entirely from the second and higher order statistical properties of the "output" signal to be modelled and can therefore be classed as blind in nature. These equations are highly nonlinear and their solution is achieved through a judicious use of reliably measured statistical features of the signal to be modelled, in conjunction with appropriate constraints and penalty functions. Examples are given to illustrate the method and it is evident from those that this novel approach is producing useful results in contexts that have been hitherto unattainable.

In this paper, we propose an iterative mixed norm image restoration algorithm. A functional which combines the least mean squares (LMS) and the least mean fourth (LMF) functionals is proposed. A function of the kurtosis is used to determine the relative importance between the LMS and the LMF functionals. An iterative algorithm is utilized for obtaining a solution and its convergence is analyzed. Experimental results demonstrate the capability of the proposed approach.

A novel Voice Activity Detector is presented that is based on Source Separation techniques applied to single sensor signals. It offers very accurate estimation of the endpoints in very low Signal to Noise ratio conditions, while maintaining low complexity. Since the procedure is totally iterative, it is suitable for use in real-time applications and is capable of operating in dynamically adapting situations. Results are presented for both White Gaussian and Car Engine background noise. The performance of the new technique is compared with that of the GSM Voice Activity Detector. 1. Introduction Voice Activity Detection (VAD) is important in many areas of speech processing technology, such as noise reduction, voice recognition, speech coding etc, and has been extensively studied ([7], [5], [1]). Most of the existing techniques focus on relatively mild noise conditions (small positive SNR, for example the conditions found in an office environment). The work presented in this paper focus...

The purpose of this contribution is to present a new approach for the prediction of speech signals that is appropriate to speech coding. The procedure is based upon the principles of blind equalisation. In an earlier publication we examined these principles from the prediction point of view as a general method. The present contribution examines the approach in relation to speech signal representation for coding and compression. The method outlined in this contribution offers significant advantages compared to the standard prediction methods. This is because the signal estimation is carried out on a sample by sample basis, it needs no estimation of the covariance matrix or some other long term statistical attributes, it makes no assumption on a minimum phase vocal tract transfer function, and hence, it is faithful to the nonstationarities in the analysed signal. The method can be seen as mapping a given signal into a set of signals of increased correlational properties, which in turn may be so mapped until the signals exhibit piecewise constant behaviour. At this stage they are easily modelled and the reverse process can be put into effect for the original signal reconstruction. Speech signals can be rendered piecewise constant approximately after some signal decompositions. Then, the percentage error of predicted value is involved to examine if the decomposition stops. Examples illustrating these principles are included.

In this paper the problem of nonlinear signal modeling is examined from a higher-order statistical perspective. The approach taken involves the use of second order Volterra kernels which are derived from a joint operation on second and third order moments of the signal. The paper describes the fundamental issues of the various components of the approach. The nonlinear equations involved are solved by means of unconstrained Lagrange programming neural networks. The results of the entire modeling scheme contained in this paper are very encouraging.

Orthogonal functions play a fundamental role in the representation of signals and are crucial in many forms of analysis and processing. The article is concerned with the discrete time Kautz (1954) extension of Laguerre functions as a basis of signal representation. There is a wealth of properties that follow from such functions that have profound implications in many areas of digital signal processing. The fundamental result is that the linear manifold of the projection of a given signal into the adopted space is obtainable by simple complex interpolation at a well defined set of points on the z-plane.>