Gait recognition is essential for early diagnosis of movement disorders. The integration of new technologies can enhance the early identification of these conditions. Many current studies use Spatio-Temporal Graph Convolutional Networks (ST-GCN) that depend on skeletal data, however, these models often require substantial memory, which limits their use in clinical settings. This work introduces an improved lightweight ST-GCN model that merges temporal convolution, weighted fusion, and GRU units to analyze movement patterns and extract spatiotemporal features for gait classification. Additionally, we present a novel Cross-Graph Relational Knowledge Distillation (CGRKD) technique that transfers both spatial and temporal relationship knowledge from a larger model to a more compact one by using shared memory and relational alignment among skeletal joints. Our CGRKD approach preserves important movement relationships between joints while lowering computational complexity, thus enhancing the model’s suitability for clinical use. Experimental results on the KOA-NM, PD-WALK, and ATAXIA datasets indicate that our method surpasses existing literature methods. The code will be available upon acceptance of the paper.

Gait recognition is essential for the early diagnosis and monitoring of movement disorders such as Knee Osteoarthritis (KOA) and Parkinson's Disease (PD). This study presents a new method for skeleton-based gait recognition. Our approach combines Spatio-Temporal Graph Convolutional Networks (STGCN) and Long Short-Term Memory (LSTM) layers to analyze movement data. The STGCN blocks capture spatial and temporal relationships between human joints, while the LSTM layers enhance the model's ability to recognize longterm gait patterns. By incorporating knowledge distillation, our method effectively transfers insights from a complex teacher model to a streamlined student model, improving both accuracy and computational efficiency. We conducted our evaluations on two public datasets for KOA and PD. The results show that our approach outperforms state-of-the-art performance, offering a reliable tool for the clinical assessment and monitoring of gaitrelated disorders.

The advancement of image restoration, especially in reconstructing missing or damaged image areas, has benefited significantly from self-supervised learning techniques, notably through the recent Masked Auto Encoder (MAE) strategy. In this project, we leverage eye-tracking data to enhance image reconstruction quality, and more specifically, with fixation-based saliency combined with the MAE strategy. By examining the emergent properties of representation learning and drawing parallels to human perceptual observation, we focus on how eye-tracking data informs the selection of image patches for reconstruction, aligning computational methods with human visual perception. Our findings reveal the potential of integrating eye-tracking insights to improve the accuracy and perceptual relevance of self-supervised learning models in computer vision. This study thus underscores the synergy between computational image restoration methods and human perception, facilitated by eye-tracking technology, opening new directions and insights for both fields. Our experiments are available for reproducibility in this GitHub Repository.

The evaluation of 3D point cloud quality is a critical component in the development of immersive multimedia systems for real-world applications. While perceptual quality evaluation technics for 2D images and videos have reached high performances, developing robust and efficient blind metrics for point cloud quality assessment is still challenging. In this paper, we propose a no-reference point cloud quality assessment method that evaluates the quality of degraded 3D objects using an end-to-end point-based multi-stream model. To capture the geometric degradation of the point cloud, we incorporate normals, curvatures and geometric coordinates. Then, we divide the distorted object into sub-objects, which are fed to a multi-stream network to extract significant features of the geometric degradation. Afterward, these features are used to predict the quality of each sub-object, and the perceptual quality score of the point cloud is obtained by averaging the quality scores of all sub-objects. Experimental results demonstrate that the proposed model achieves promising performance compared to state-of-the- art full and reduced methods.

In this paper, we propose a novel method of explainable deep learning for COVID-19 detection in Chest X-ray (CXR) images, employing a Texture-Shape approach. The method extracts texture information from the first convolution layer and shape features from the last convolution layer of Deep Neural Networks, namely CNN, VGG16, and ResNet50. To generate the final explainable map, the extracted texture and shape heatmaps are fused using a guided filter-based method. Our approach is versatile and can be adapted to work with classical explainability methods commonly used in the literature. We validate the efficacy of our method on a well-known COVID-19 database comprising CXR images. We conduct extensive experiments to assess its performance against classical explainability methods such as FEM (Feature Extraction Map) and GradCam (Gradientweighted Class Activation Mapping). For evaluation, we calculate metrics such as the average drop and increase in confidence scores. The obtained results demonstrate that fusing texture and shape information leads to significantly improved explainability compared to conventional methods.

This paper proposes a novel method for colored mesh visual quality assessment based on graph features learning. We first extract color features from each distorted colored mesh data. Then, we map them into a weighted graph with weights derived from the color features. One then computes various local topological properties of the network nodes (Degree, Strength, Clustering coefficient). Estimates of their statistical properties (Mean, Variance, Skewness, Kurtosis, Entropy) form a signature vector are used by a machine learning algorithm. The random forest regression is used to predict the quality score. Experiments are conducted on the publicly available CMDM database specifically constructed for the colored mesh quality assessment task. The proposed method is compared to the most influential and effective full and no-reference methods (CMDM and NR-NSS). The excellent correlations with subjective decisions prove its good performance.

This paper deals with a saliency-based no-reference (NR) method for 3D point cloud (PC) quality assessment. For this purpose, we firstly compute 3D visual saliency map for each distorted point cloud. Then, we use a threshold-based filter to select the most salient points. From these, we extract both geometrical a perceptual attributes. Estimates of their statistical properties (Entropy, Standard deviation, Skewness, Kurtosis, Median and Mean) form a features vector. In the end, the Support vector regressor (SVR) is utilized for the characteristics regression and the quality score prediction. To validate our method, a set of experiments are conducted on an open subjective colored point cloud dataset (SJTU-PCQA). Results show that the suggested method exceeds some competing methods accord-ina to correlation with average opinion score.

This paper addresses the problem of blind stereoscopic image quality assessment (NR-SIQA) using a new multi-task deep learning based-method. In the field of stereoscopic vision, the information is fairly distributed between the left and right views as well as the binocular phenomenon. In this work, we propose to integrate these characteristics to estimate the quality of stereoscopic images without reference through a convolutional neural network. Our method is based on two main tasks: the first task predicts naturalness analysis based features adapted to stereo images, while the second task predicts the quality of such images. The former, so-called auxiliary task, aims to find more robust and relevant features to improve the quality prediction. To do this, we compute naturalness-based features using a Natural Scene Statistics (NSS) model in the complex wavelet domain. It allows to capture the statistical dependency between pairs of the stereoscopic images. Experiments are conducted on the well known LIVE PHASE I and LIVE PHASE II databases. The results obtained show the relevance of our method when comparing with those of the state-of-the-art. Our code is available online on https://github.com/Bourbia-Salima/multitask-cnn-nrsiqa_2021

We propose in this work a graph-based deep learning method for mesh visual quality assessment. To carry out this proposal, we transform a given distorted mesh to a graph represented by its adjacency matrix, and extract a set of geometric and perceptual features to be stored into a feature matrix. The two matrices are then learned to a graph convolutional network (GCN). The network is composed by two convolutional layers followed by a max-pooling layer. The Softmax classifier is used to predict the quality relying on the node classification problem. Five classes are considered according to the ground truth scores: very bad, bad, medium, good or excellent quality. Experiments are conducted on two publicly available databases specifically constructed for the quality assessment task. Our method is compared to some influential and effective full and reduced reference methods. The good performance is proven by the excellent correlations with subjective decisions.

We propose in this paper a novel objective method to evaluate the perceived visual quality of 3D meshes. The proposed method in no-reference, it relies only on the distorted mesh for the quality estimation. It is based on a pre-trained convolutional neural network (i.e VGG to extract features from the distorted mesh) and handcrafted features extracted directly from the 3D mesh (i.e curvature and dihedral angle). A General Regression Neural Network (GRNN) is used to learn the statistical parameters of the feature vectors and estimate the quality score. Experimental results from for subjective databases (LIRIS masking, LIRIS/EPFL generalpurpose, UWB compression and LEETA simplification) and comparisons with objective metrics cited in the state-of-the-art demonstrate the efficacy of the proposed metric in terms of the correlation to the mean opinion scores across these databases.

In this paper, we leverage the properties of non-Euclidean Geometry to define the Geodesic distance (GD) on the space of statistical manifolds. The Geodesic distance is a real and intuitive similarity measure that is a good alternative to the purely statistical and extensively used Kullback-Leibler divergence (KLD). Despite the effectiveness of the GD, a closed-form does not exist for many manifolds, since the geodesic equations are hard to solve. This explains that the major studies have been content to use numerical approximations. Nevertheless, most of those do not take account of the manifold properties, which leads to a loss of information and thus to low performances. We propose an approximation of the Geodesic distance through a graph-based method. This latter permits to well represent the structure of the statistical manifold, and respects its geometrical properties. Our main aim is to compare the graph-based approximation to the state of the art approximations. Thus, the proposed approach is evaluated for two statistical manifolds, namely the Weibull manifold and the Gamma manifold, considering the Content-Based Texture Retrieval application on different databases.

Autism spectrum disorder (ASD) is a neurodevelopmental disorder that touches children in an early age and alters the function of the brain. Previous studies put forward theories of under and over-connectivity between regions of the autistic brain. Hence, to understand the disorder and find an early diagnosis corroborating the existing theories is of central importance. In this paper, we propose a framework that takes into account the properties of over-connectivity in the autistic brain using the maximum spanning tree (MaxST), since this latter is known to describe high connectivity values. The novelty of the proposed approach is to adopt elimination of the information related to the overconnectivity theory, i.e elimination of the MaxST. This permits to measure the impact of the suppression and thus to well emerge the aforementioned connectivity alterations. With an overall objective of facilitating the early diagnosis of this disorder. The tested dataset is the large multi-site Autism Brain Imaging Data Exchange (ABIDE). The results show that this approach provides accurate prediction up to 70%. They also highlight the importance of every parameter used in all the steps that lead to the final result.

OsteoArthritis (OA) is the most common disorder of the musculoskeletal system and the major cause of reduced mobility among seniors. The visual evaluation of OA still suffers from subjectivity. Recently, Computer-Aided Diagnosis (CAD) systems based on learning methods showed potential for improving knee OA diagnostic accuracy. However, learning discriminative properties can be a challenging task, particularly when dealing with complex data such as X-ray images, typically used for knee OA diagnosis. In this paper, we introduce a Discriminative Regularized Auto Encoder (DRAE) that allows to learn both relevant and discriminative properties that improve the classification performance. More specifically, a penalty term, called discriminative loss is combined with the standard Auto-Encoder training criterion. This additional term aims to force the learned representation to contain discriminative information. Our experimental results on data from the public multicenter OsteoArthritis Initiative (OAI) show that the developed method presents potential results for early knee OA detection.

Texture is a fundamental characteristic of many types of images, especially those with significant rotation, scale illumination, and viewpoint change. Texture image classification is one of the challenging problems that have various applications such as remote sensing, material recognition, and computer-aided medical diagnosis, etc. Various Computer vision techniques have been used. More recently, Deep learning architectures demonstrated impressive results. This paper aims to investigate combining two feature extraction methods: Handcrafted-based and CNN-based in a two-stream neural network architecture. We believe that Statistical features could enhance the performance of the CNN architecture, especially in the case of small datasets. To test our approach we used two challenging datasets, the Describable Textures Dataset (DTD) and Flicker Material Database (FMD). Results showed that our two-stream neural network which has an image as a first stream and a statistical feature vector as a second stream achieve better results than a Convolutional neural network achieved with just the RGB image as input. The Xception network [9] combined with SIFT-FV demonstrated an accuracy superiority for both datasets.

In this work, we propose a convolutional neural network (CNN) framework to estimate the perceived visual quality of 3D meshes without having access to the reference. The proposed CNN architecture is fed by small patches selected carefully according to their level of saliency. To do so, the visual saliency of the 3D mesh is computed, then we render 2D projections from the 3D mesh and its corresponding 3D saliency map. Afterward, the obtained views are split to obtain 2D small patches that pass through a saliency filter to select the most relevant patches. Experiments are conducted on two MVQ assessment databases, and the results show that the trained CNN achieves good rates in terms of correlation with human judgment.

This paper proposes a new target recognition method for inverse synthetic aperture radar (ISAR) images. This method is based on joint statistical modeling of the complex wavelet coefficients for ISAR image characterization and the sparse representation based classification (SRC) for the recognition. To extract features from an ISAR image, we first transform it in the complex wavelet domain using the dual-tree complex wavelet transform (DT-CWT). Then, we compute magnitude information for each complex subband. After that, we propose a joint statistical model for magnitude distribution, that takes into account the dependences between different orientations and scales. To do so, we adopt the copula as a multivariate model thanks to its suitability to capture jointly the subband marginal distribution and the dependence structure. For the recognition step, we exploit SRC which recovers the test descriptor to classify over a given dictionary composed by the training descriptors. This method classifies the test sample as the class whose training samples can generate the minimum sparse representation error. Experimental results on ISAR images database show that using copula and sparse classifier improve significantly the recognition rates compared to classical models and classifiers.

In this paper, we propose a new method for blind mesh visual quality assessment using a deep learning approach. To do this, we first extract visual representative features by computing locally curvature and dihedral angles from each distorted mesh. Then, we determine from these features a set of 2D patches which are learned to a convolutional neural network (CNN). The network consists of two convolutional layers with two max-pooling layers. Then, a multilayer perceptron (MLP) with two fully connected layers is integrated to summarize the learned representation into an output node. With this network structure, feature learning and regression are used to predict the quality score of a given distorted mesh without needing to a reference mesh. Experiments are conducted on LIRIS masking and the general-purpose databases and results show that the trained CNN achieves good rates in terms of correlation with human visual judgment scores.

Color and texture have been proven to be very discriminant attributes in image analysis across many works. This paper proposes a color texture analysis method based on the graph theory, in which we convert the texture in question into an undirected weighted graph and explore the shortest paths between four pairs of pixels according to different scales and orientations of the image. Basically, we extend two previously introduced approaches that consider the RGB color space, to textures in the HSV color space taking into account the nature of correlations between its color channels. In order to evaluate its performance, we applied this procedure to USPTex textures dataset. The best classification results using the standard parameters of the method are 92.25%, 91.75% and 85.72% of Accuracy (percentage of samples correctly classified). Which proves the efficiency of the proposed method compared to the results achieved by traditional methods found in literature.

Osteoporosis is one of the most common bone diseases over the world, this disease causes a high risk of fracture, an early diagnosis of the osteoporosis can help avoiding the risk of fracture. Several methods have been proposed to detect the presence of this disease. In this paper we propose a method of osteoporosis diagnosis based on bone X-Ray images, using the frequency separation and fractional Brownian motion. The classification is performed using the Support Vector Machines classifier. The proposed method shows its good classification performance achieving an accuracy rate around 94 %.

In this letter, we present a novel generic approach for radar automatic target recognition in either inverse synthetic aperture radar (ISAR) or synthetic aperture radar (SAR) images. For this purpose, the radar image is described by a statistical modeling in the complex wavelet domain. Thus, the radar image is transformed into a complex wavelet domain using the dual-tree complex wavelet transform. Afterward, the magnitudes of the complex sub-bands are modeled by Weibull or Gamma distributions. The estimated parameters of these models are stacked together to create a statistical dictionary in training step. For the recognition task, we use the weighted sparse representation-based classification method that captures the linearity and locality information of image features. In this context, we propose to use the Kullback-Leibler divergence between the parametric statistical models of training and test sets in order to assign a weight for each training sample. Experiments conducted on both ISAR and SAR images' databases demonstrate that the proposed approach leads to an improvement in the recognition rate.

Osteoporosis diagnosis has attracted particular attention in recent decades. Textured images from the microarchitecture of osteoporotic and healthy subjects show a high degree of similarity, increasing the difficulty of classifying such textures. Thus, the evaluation of osteoporosis from the bone X-ray images presents a major challenge for pattern recognition and medical applications. The purpose of this paper is to use the fractional Brownian motion (fBm) model and the probability density function of its increments to compute a similarity measure with the Rao geodesic distance to classify trabecular bone X-ray images. When evaluated on synthetic fBm images (test vectors) with the well-known Hurst parameter H, the proposed method met our expectations in which a good classification of the synthetic images was achieved. A clinical study was conducted on textured bone X-ray images from two different female populations of osteoporotic patients (fracture cases) and control subjects. Using the proposed method, an area under curve rate of 97% was achieved.

This paper deals with a new anisotropic discrete dual-tree wavelet transform (ADDTWT) to characterize the anisotropy of bone texture. More specifically, we propose to extend the conventional discrete dual-tree wavelet transform (DDTWT) by using the anisotropic basis functions associated with the hyperbolic wavelet transform instead of isotropic spectrum supports. A texture classification framework is adopted to assess the performance of the proposed transform. The generalized Gaussian distribution is used to model the distribution of the sub-band coefficients. The estimated vector of parameters for each image is then used as input for the support vector machine classifier. Experiments were conducted on synthesized anisotropic fractional Brownian motion fields and on a real database composed of osteoporotic patients and control cases. Results show that the ADDTWT outperforms most of the competing anisotropic transforms with an area under curve rate of 93%.

The work presented in this paper is part of the filed of automatic recognition of radar targets. Thus, for assistance in target recognition, we propose a new approach to extract efficient feature from synthetic aperture radar (SAR) images. The proposed approach deals with a combination of two feature descriptors obtained from two methods. In the first method, we perform the dual-tree complex wavelet transform (DT-CWT) on SAR image, and then, the complex subbands magnitudes are modeled by a non-Gaussian statistical model. In the second method, we use the scale invariant feature transform (SIFT). Due to the fact that SIFT descriptor is limited to a huge dimension, we propose to model its statistical behavior using a non-Gaussian statistical model in order to overcome this limit. The combination of the resulting Weibull or Gamma statistical parameters for the both DT-CWT and SIFT methods are selected as a feature vector. To validate our appraoch, the classification results are provided using Polynomial kernel based support vector machines (SVM) classifier. The experimental results using SAR images database show the benefits of the proposed approach to extract feature descriptor.

The importance of phase features for texture analysis has been earlier established for many image processing applications. However, the modeling of the phase data faces some difficulties as its information gathers data with rotating values and thus highly sensitive to distortions. Motivated by its ability to capture different shapes of histograms, in this communication, we propose the Gaussian Mixture Model (GMM) to characterize the behavior of relative phase. The Maximum-likelihood Estimator (MLE) is used to estimate the GMM parameters. To investigate the relevance of the GMM model for relative phase data, a feature vector incorporating the estimated parameters is proposed for a multiclass classification task. Experiments are conducted on textures from VisTex and Brodatz databases. Results demonstrate that the GMM model fit well relative phase data. In addition, higher rates of accuracy, precision and recall, 95.35%, 95.50% and 95.40%, respectively, were achieved for Brodatz textures using the proposed feature vector. This suggests the potential usefulness of the probabilistic proprieties for texture analysis.

This paper presents a univariate multi-model for color texture characterization in luminance-chrominance (LC) color spaces. Such color spaces are known to separate between luminance and chrominance, contrary to RGB representation. The magnitude of the luminance component is then represented via the Gamma distribution. For chrominance, a new proposed information is investigated, namely the extended relative phase (ERP). This latter allows to capture the phase information inter-chrominance components. This information is characterized using the circular Wrapped cauchy (WC) distribution. The resulting {Gamma, WC} model has the advantage to be a linear-circular model. It, thus, represents color components considering their nature. Secondly, it permits to catch an information of dependence between chrominance components while being univariate. This means a modeling with simple feature extraction and straightforward implementation of similarity measurement, compared to the multivariate characterization widely used for color textures. Results on the Vistex database show that the study of {magnitude, ERP} information in LC color spaces, improves the retrieval performances compared to the RGB based characterization.

This paper deals with the classification and segmentation of seafloor images recorded by sidescan sonar. To address this problem, related to texture analysis, a supervised approach is considered. The features of the textured images are extract by characterizing the wavelet coefficients through parametric probabilistic models. In this contribution, the generalized Gaussian distribution and the α-stable distribution are used. For the classification step, two classifiers are considered: the k-nearest neighbor algorithm, that exploit the Kullback-Leibler divergence as similarity measurement, and the support vector machines. Experimental results on sonar images demonstrate the effectiveness of the proposed approach for sonar image classification and segmentation.

Image segmentation has an important role in many image processing applications. Several methods exist for segmenting an image. However, this technique is still a relatively open topic for which various research works are regularly presented. With the recent developments on complex networks theory, image segmentation techniques based on graphs has considerably improved. In this paper, we present a new perspective of image segmentation, by applying three of the most efficient community detection algorithms, Louvain, infomap and stability optimization based on the louvain algorithm, and we extract communities in which the highest modularity feature is achieved. After we show that this measure is invariant to non-structural change on image, which mean that the image segmentation is also invariant to rotation. Finally we evaluate the three proposed algorithms for Berkeley database images, and we show that our results can outperform other segmentation methods in terms of accuracy and can achieve much better segmentation results.

In this paper, we address the problem of rotation invariance in the context of texture retrieval. For this, we propose a framework based on the well-known copula theory which is considered one of the most powerful statistical tools. Prior to apply a such model, we first use the steerable pyramid SP as one of the most relevant transforms. Then, we build a steerable Gaussian copula model which offers a good fitting of the SP coefficients distribution while taking into consideration their rotation invariance property. Finally, we derive a closed-form of the Jefferey divergence as a similarity measure. The latter consists on an angular alignment between the query and the target texture features. Experiments have been conducted on USC database, good performances in term of retrieval rates are achieved compared to previously proposed copula models.

In this paper, we propose a novel feature extraction scheme based on the multi-resolution curvelet transform for face recognition. The obtained curvelet coefficients act as the feature set for classification, and are used to train the ensemble-based discriminant learning approach, capable of taking advantage of both the boosting and LDA (BLDA) techniques. The proposed method CV-BLDA has been extensively assessed using different databases: the ATT, YALE and FERET, Tests indicate that using curvelet-based features significantly improves the accuracy compared to standard face recognition algorithms and other multi-resolution based approaches.

In this paper, an efficient local appearance feature extraction method based steerable pyramid (S-P) is proposed for face recognition. Local information is extracted from S-P sub-bands using block-based statistics. The underlying statistics allow us to reduce the required amount of data to be stored. The obtained local features are combined at the feature and decision level to enhance face recognition performance. Experimental results on ORL, Yale and FERET face databases convince us that the proposed method provides a better representation of the class information and obtains much higher recognition accuracies.

In this paper, a new spatiotemporal filtering scheme is described for noise reduction in video sequences. For this purpose, the scheme processes each group of three consecutive sequence frames in two steps: 1) estimate motion between frames and 2) use motion vectors to get the final denoised current frame. A family of adaptive spatiotemporal L-filters is applied. A recursive implementation of these filters is used and compared with its nonrecursive counterpart. The motion trajectories are obtained recursively by a region-recursive estimation method. Both motion parameters and filter weights are computed by minimizing the kurtosis of error instead of mean squared error. Using the kurtosis in the algorithms adaptation is appropriate in the presence of mixed and impulsive noises. The filter performance is evaluated by considering different types of video sequences. Simulations show marked improvement in visual quality and SNRI measures cost as well as compared to those reported in literature.

In this paper, a new multichannel spatio-temporal filter is described for color video sequences restoration with the presence of non-Gaussian zero-mean additive noise. To address this problem, we propose a multichannel L-filter optimized by the least mean Kurtosis (LMK) algorithm. Prior to filtering, motion compensation is performed by a robust simultaneous estimation for all color components using an affine linear model. Then, we apply the proposed filter to the reconstructed frames. Experiments were performed on real color video sequences, and performance comparisons have been made both in RGB and L*a*b* color spaces.

This paper presents a 3-D adaptive Volterra filter with an LMS type adaptation algorithm. This filter is used for equalizing an unknown channel with some point-wise nonlinearity and restoring image sequences degraded by this channel. First a robust higher order statistics (HOS) based motion estimation method is employed. Then, a 3-D Volterra equalizer is applied. The performance of this algorithm is demonstrated using real image sequences.