Abstract Electrical Impedance Tomography (EIT) as a non-invasive imaging method, faces two main challenges due to the inherent difficulty in solving the inverse image reconstruction problem. The main difficulty lies in simultaneously achieving two key goals: Accurately estimating conductivity values and clearly reproducing structural boundaries. Existing methods typically excel at one aspect while compromising the other. In this work, we present an advanced ensemble learning approach combining complementary deep learning models to overcome these challenges. The method integrates two specialized ensembles: A gradient-boosting framework with an enhanced 1D-CNN-GRU for conductivity prediction and a Dense Attention Network (DA-Net) for boundary shape reconstruction, unified through a stacking ensemble with a custom loss function optimizing both conductivity accuracy and structural preservation by combining Rooted Mean Squared Error (RMSE) for quantitative accuracy and Image Correlation Coefficient (ICC) in a weighted formulation. The model is first validated in a well-controlled water tank setting. Simulation results show an ICC of 0.975 and Relative Image Error (RIE) of 0.171. Experimental validation further confirms the method’s robustness, maintaining consistent reconstruction quality for different object positions and materials. The method was then evaluated using a more complex, simulated lung model that includes realistic anatomical features, changes in conductivity distribution, and tissue anomalies. Results demonstrate improved performance compared to baseline methods by an average 2.67% increase in ICC and 24.55% reduction in RIE. The results confirm that the novel systematic ensemble approach successfully addresses the dual challenges of conductivity prediction and boundary preservation for effective anomaly detection in EIT imaging.

Parkinson’s disease (PD) is a chronic and progressive neurodegenerative disorder that severely impacts motor functions, including handwriting. Early and precise detection is essential for timely intervention and effective treatment. In this paper, we propose a novel hybrid approach that integrates handcrafted and deep learning-based (DL) features extracted from handwriting samples. By leveraging both feature types, our method provides a more comprehensive analysis, enhancing the accuracy and robustness of PD detection. Indeed, Handcrafted features provide explicit, interpretable descriptors that capture domain-specific patterns while DL features encode high-level abstract representations. We first extracted DL features using a fine-tuned transfer learning model based on ResNet50 and combined them with a set of handcrafted features. Similarly, DL features were obtained using a second fine-tuned transfer learning model based on MobileNetV2, and the derived features were integrated with the same handcrafted feature set for further analysis. Afterward, a feature selection step is performed.The conducted experiments exploiting a voting classifier, demonstrates that the proposed hybrid approach achieve promising results reaching an accuracy of 91.51%.

Accurate State of Charge (SoC) estimation is essential for the efficient management of Li-ion batteries, especially under varying operating conditions. Traditional filtering approaches rely on predefined battery models, requiring prior knowledge of internal dynamics, which may introduce inaccuracies in real-world applications. In this study, we propose a data-driven SoC estimation method based on a Bidirectional Long Short-Term Memory (Bi-LSTM) network enhanced with a self-attention mechanism, designed to identify and prioritize key time steps in the battery’s charge-discharge cycle. The model takes voltage, current, and temperature as inputs and has been validated across multiple battery profiles under diverse temperature conditions. Experimental results demonstrate that the proposed approach achieves an estimation accuracy of approximately 98%, with a Root Mean Squared Error (RMSE) below 1.7, significantly improving the reliability of SoC predictions. By using both past and future states, along with attention-driven feature weighting, the proposed model enhances SoC estimation robustness across different operating scenarios.

Parkinson’s disease (PD) is a progressive neurodegenerative disorder that leads to motor impairments. Early detection is crucial for effective treatment; however, conventional diagnostic methods are often costly, time-consuming, and inaccessible, limiting their widespread clinical adoption. To address these challenges, we propose a novel hybrid approach that merges Deep Learning (DL) features extrated from handwriting images with voice characteristics using an optimized machine learning (ML) classification technique. The integration of multimodal data enhances robustness by reducing dependency on a single biomarker, making PD diagnosis more reliable. We begin by augmenting both of the datasets size to expand samples diversity. Then, we combine DenseNet201’s detailed features with ResNet50’s robust spatial features to enhance analytical precision and capture both fine-grained and high-level patterns of handwriting images. The obtained DL features are fused with voice characteristics leveraging complementary information. Afterwards, feature selection is performed using Fisher’s score to retain the most relevant attributes, further boosting classification accuracy. We achieve an accuracy of 92.31%, demonstrating superior performance compared to state-of-the-art methods.

Keystroke dynamics, a behavioral biometric modality, offers promising applications in authentication and intrusion detection systems. However, the scarcity of publicly available datasets due to privacy concerns limits research progress. This paper presents a Generative Adversarial Network (GAN) framework to generate synthetic keystroke dynamics data that closely mimics real-world patterns. Using the DSL-StrongPasswordData dataset, we pre-process and normalize timing features and train a GAN with 100-dimensional latent space, LeakyReLU activations, and binary cross-entropy loss. We evaluated the synthetic data through visual comparisons (boxplots, t-SNE projections) and statistical tests (Kolmogorov-Smirnov), demonstrating that the generated distributions align with real data (p-value > 0.05 for key features). Our results highlight the potential of the GAN for sharing data that preserve privacy and increase training sets for keystroke-based models.

Image forgery detection is crucial in digital forensics, cybersecurity, and legal investigations. Despite advancement, detecting subtle manipulations like copy-move forgeries remains challenging due to increasingly realistic images. This paper proposes a hybrid approach that combines a pretrained VGG16 model for feature extraction, cosine similarity for block-level comparison, and support vector machines for classification, addressing key limitations of existing methods. Through a comparative evaluation of CNN architectures, VGG16 is identified as the most effective for extracting discriminative features in this context. Cosine similarity quantifies the similarity between image block features to enable the model to focus more effectively on the tampered regions that closely resemble the original ones, and SVMs are leveraged to classify authentic versus forged regions. This novel integration of deep learning for feature extraction and classical classification techniques is highly accurate with minimal false positives, without relying on handcrafted features. Experiments on the MICC-F2000 dataset demonstrate the method’s strong performance, achieving 99.59% precision, 98.00% recall, and a 0.99 F1-score.

For a few decades, automatic license plate recognition (ALPR) has been considered as a powerful intelligent video analytics tool to deal with the growing number of vehicles worldwide. Most recent ALPR landmark achievements are contingent on access to diverse and rich datasets for model training and evaluation. In this context, we present a naturalistic dataset of European license plates, named ELP 1.0. Unique to ELP 1.0 is its vast geographical coverage, comprising a wide spectrum of European countries (20) and vehicle categories (10). Using the proposed dataset, we benchmark the performances of several baseline deep learning models, assessing their ability to detect and recognize license plates. Further, to picture the ELP 1.0 dataset’s features, we propose a try-one-subset-out setup and evaluate the generalization abilities of ALPR approaches across an array of scenarios. The experimental findings highlight ELP 1.0 challenges as well as its usefulness to question ALPR models’ capacities. The ELP 1.0 dataset, with its comprehensive annotations, features, and benchmark results, will serve as a testbed to support future ALPR research and, more broadly, intelligent transportation industrial solutions. The ELP 1.0 dataset is accessible upon request via this URL.

Ensuring the safety of workers in hazardous environments is crucial, given the potential severity of accidents. Indeed, Personal Protective Equipment (PPE), including safety helmets, vests, and boots, is vital in mitigating risks. However, adherence to PPE usage often falls short due to inadequate safety awareness among workers. To address this challenge, we aim to develop a robust safety monitoring system by leveraging advancements in deep learning and object detection technology. Our proposed method uses YOLOv7 and YOLOv8 algorithms for real-time detection and classification of safety equipment, including helmets, vests, and boots. For this aim, we carefully collect and annotate datasets from diverse sources like Roboflow and Kaggle. Then, we accurately evaluate performance metrics such as precision, recall, and speed through rigorous training and testing of various YOLOv7 and YOLOv8 configurations on annotated data batches. By conducting thorough comparative analyses, we aim to identify the most effective model for each PPE category, considering both detection accuracy and computational efficiency. Finally, the object detection system is deployed through a web interface to determine whether a person is protected or not.

State of charge (SoC) estimation is a critical aspect of managing lithium-ion (Li-ion) batteries in electric vehicles (EVs). Various Artificial Intelligence (AI) techniques have been used to enhance SoC estimation accuracy, such as convolutional neural networks, recurrent neural networks, generative networks, and more. This article provides a thorough survey of AI-based models utilized for SoC estimation in Li-ion batteries, synthesizing findings from diverse studies. By comparing the performance of different models : classical Machine Learning, convolutional, recurrent and generative, insights into the effectiveness and limitations of various AI approaches are highlighted. The review aims to guide future research efforts toward developing robust and accurate SoC estimation methods crucial for optimizing EV battery management systems.

Automatic license plate recognition (ALPR) is a powerful tool for analyzing the growing number of vehicles in major cities worldwide. However, building datasets that accurately represent vehicles and license plates (LPs) is challenging, as most current datasets are specific to certain countries and do not provide a comprehensive representation. This remains particularly true for French LPs, where there is currently no publicly available dataset that covers a wide range of scenarios such as plate inclinations, varying lighting conditions, and changing weather. To address this issue, we present a benchmark suite called PGFLP 1.0, which includes more than 14k fully annotated LPs from several vehicle categories (e.g. cars, trucks, motorcycles). PGFLP 1.0 includes three annotation levels: rectangular boxes for the LP, four vertices of the LP region, and labels for the full LP sequence. Using PGFLP 1.0, we provide an experimental evaluation of the most recent ALPR approaches, assessing their ability to detect and recognize LPs. This work provides a useful benchmark in terms of accuracy and speed for current and future research efforts in ALPR.

This paper presents a novel approach for degraded face recognition using Vision Transformer (ViT) architectures. The process begins by inputting occluded face images into a Transformer encoder, which extracts discriminative features and creates an embedding vector essential for the de-occlusion process. In the next step, a Transformer decoder refines this representation by incorporating non-occluded features from the ground truth image, employing self-attention mechanisms to integrate information from both the embedding vector and ground truth features. The decoder outputs a comprehensive image of the occluded face, enriched with details from the input and ground truth images. Finally, the decoder generates a non-occluded face image, accurately reconstructing the occluded features. ViT in both encoder and decoder stages ensures efficient extraction and refinement of facial information, balancing memory and time constraints for optimal performance even in severe degradation conditions. The study uses two publicly available datasets, EKFD and IST-EURECOM LFFD. It evaluates the effectiveness of face recognition algorithms under various levels of face degradation such as partial occlusions, lighting variations, head poses, and facial expression changes.

MobileNet is a lightweight convolutional neural network optimized for resource-limited environments. However, its use of depthwise separable convolutions to minimize parameters and computation can reduce accuracy due to oversimplified channel interactions. Principal Component Analysis (PCA) can address this issue by reducing the dimensionality of weight matrices while preserving key features. Applying PCA can help maintain accuracy and compress the model simultaneously. Based on this, we propose a novel copy-move forgery detection approach based on MobilNetV2 and PCA for feature extraction, and a random forest for classification. This allows us to enhance MobileNet accuracy while keeping its model size compact. The results of our experiments conducted on the MICC-F2000 dataset reveal that the proposed hybrid lightweight model outperforms the individual transfer learning structures and the existing literature, achieving 96.37% accuracy.

With the advancement and widespread use of digital devices, capturing images has become effortless in any location. Images serve as evidence, making the authenticity of digital images increasingly critical. Some individuals alter images by adding or removing elements, rendering the images unreliable. Consequently, detecting image forgery has become essential. The evolution of image editing software has intensified this issue within the realm of computer vision. Recently, a variety of algorithms have been developed to identify image forgery. However, with the progress in digital technology, the ease of image manipulation has led to a surge in forgery cases, presenting significant obstacles when verifying their authenticity. Thus, there is a pressing requirement for effective forgery detection methods. This paper introduces a novel copy-move forgery detection approach based on the fusion of a handcrafted forgery detection system utiliszing the scale-invariant feature transform and the support vector machine with a deep-learning-based forgery detection system using VGG16. The aim is to address the challenges of subtle forgery detection and blending and seamless Integration. We conduct the experiments on the MICC_F2000 image manipulation dataset and assess the efficacy of the proposed approach, achieving $\mathbf{9 6 . 7 5 \%}$ accuracy, $\mathbf{1 0 0 \%}$ precision and $95.5 \%$ F1-score. This research demonstrates superior performance compared to state-of-the-art methods.

The importance of ensuring the authenticity and reliability of digital images has grown significantly, primarily due to the ease of modifying such images with the progress of digital image editing tools. Consequently, there is a growing emphasis on the development of techniques for detecting image manipulation. One particular area of focus in digital image authentication is copy-move forgery detection. This paper presents a survey and a comparative study on copy-move forgery detection techniques in digital images, databases, and evaluation metrics. The study aims to provide insights into the effectiveness of different methods in detecting copy-move forgeries. The paper discusses prominent detection techniques, including block-based, keypoint-based, transform domain, hybrid methods, deep learning, and GAN approaches. The findings highlight the strengths, weaknesses, and key similarities and differences among the approaches. This study contributes to the understanding of the state-of-the-art in copy-move forgery detection and provides guidance for future research in this field.

Automatic analysis of histopathological images is a very challenging topic and it involves various tasks. While interpreting Whole Slide Images (WSI), pathologists require advanced tools to perform their analysis. Thus, we propose a content-based histopathological image retrieval (CBHIR) based on incremental phenotyping approach. The CBHIR aims to assist practitioners in the decision making into a diagnosis process while providing similar cases which clinicians can refer to interpret a current one. The phenotyping approach consists mainly in delving into the large histological image in order to discover the correlations that exist between the different regions. The phenotyping graph will be applied in the CBHIR system.

The use of biometric technology in authentication and integrity verification systems necessarily gives rise to issues relating to the security and privacy of transmitted and stored templates. Many techniques have been proposed in the literature involving crypto-biometric algorithm, template transformation method and watermarking reinforcement scheme. Watermarking reinforcement relying on biometrics usually combines the watermarking technique with a feature transformation scheme. We present a new watermarking reinforcement scheme including two security levels to protect biometric data. The first security level is proposed to verify the integrity of the fingerprint while transmitted or stored based on a watermarking approach. To avoid any registration of the stored biometric template, a second security level is developed based on cancellable transformation. The watermarking of the fingerprint image is applied in the wavelet packet decomposition multiresolution domain using a binary watermark derived from the fingerprint minutiae. The developed transformation is designed to explore the relative relation between minutiae in the pair-polar coordinate domain. The transformed minutiae satisfy the non-invertibility recovery of the original ones. Under various scenarios, the suggested scheme is evaluated using the public fingerprint database BioSecure and FVC2002 DB1. The realized testing and derived results show the robustness of the first security level under various attacks and the satisfaction of the second level security to protect biometrics template with an insignificant degradation of the authentication performance.

By adopting keystroke dynamics, authentication applications can integrate advanced identity proofing technology for detecting fraud and prevent unauthorized access. However, understanding a user's keystroke dynamics behavior in real applications is a challenging task regarding that this behavior is notably changing over time. To mitigate this problem, we apply, in this paper, the long short-term memory (LSTM) model that recognizes a continuous sequences of keystroke dynamics to identify users of public datasets. We also consider the bidirectional long short-term memory (BLSTM) as it maintain information about the future data. Hence, collecting information about intra-class variations of the keystroke dynamics from both past and future data, is an interesting solution to our problem. The obtained results are promising since we obtained an accuracy rate over than 60% for both architectures when dealing with public databases.

Autonomous vehicles present a promising opportunity in the future of transportation systems by providing road safety. As significant progress has been made in the automatic environment perception, the detection of road obstacles remains a major challenge. Thus, to achieve reliable obstacle detection, several sensors have been employed. For short ranges, the Ultra-Wide Band (UWB) radar is utilized in order to detect objects in the near field. However, the main challenge appears in distinguishing the real target’s signature from noise in the received UWB signals. In this paper, we propose a novel framework that exploits Recurrent Neural Networks (RNNs) with UWB signals for multiple road obstacle detection. Features are extracted from the time-frequency domain using the discrete wavelet transform and are forwarded to the Long short-term memory (LSTM) network. We evaluate our approach on the OLIMP dataset which includes various driving situations with complex environment and targets from several classes. The obtained results show that the LSTM-based system outperforms the other implemented related techniques in terms of obstacle detection.

Terrorism remains to be among the most significant global dangers by 2020. A considerable literature has grown around the theme of the fight against terrorism. However, to the best of our knowledge, no research work has been performed on terrorist face images caught in real-world and uncontrolled conditions. In this context, we propose in this paper a terrorist suspect face recognition database. Following substantial studies and analysis on some terrorist attacks that have occurred since 2013, faces of terrorists have been collected from the net. The LATIS-PACTE-PROFILER database is made available to the research community. For validation and benchmarking purposes, we propose a face identification approach based on the HOG features and the SVM classifier.

Biometrics has for objective to identify or verify the identity of an individual based on morphological or behavioral characteristics. A biometric system can be attacked by presenting a biometric data to the capture subsystem with the goal of interfering it, that is called a presentation attack. Covid, panther, shadow monster and dragon are the investigated presentation attacks associated to the Doddington Zoo Menagerie (which classify users in different categories considering their performance behavior when using biometric systems). In this work, we examined the robustness of each genuine class of the biometric menagerie against the proposed presentation attacks. The achieved experiments are applied to the keystroke dynamics modality. Owing to the adaptive strategy, we depicted each genuine category that is most vulnerable to a specific presentation attack class. We find that the impact of covid, panther, shadow monster and dragon attempts are more pronounced when compared to chameleons, worms, doves and phantoms classes respectively. The obtained results, point out that adding imposter labels to Doddington zoo may lead to a better assessment of biometric authentication systems and promotes the interpretation of their performances.

3D data are an important resource for many computer-based applications, as they provide valuable depth cues about the full geometry of 3D associated objects. They become even more valuable as regards 3D face/ facial expression recognition using deep learning. Indeed, two main challenges remain under study. The first is how to resume 3D faces with a discriminative representation from a 3D point cloud while exploiting an adequate Deep Neural Network (DNN). The second is the lack of large 3D facial datasets. To address the first issue, we propose to exploit solely geometric information while applying DNN. Hence, in order to deal with high resolution face scans with a rich point cloud representation, we extract point-based representations using various sampling strategies. Different keypoint sets are used, ranging from a small set of points of interest (i.e. landmarks) to point sets sampled from a curve-based representation, as well as scale-invariant feature transform keypoints. As for the second issue and in order to overcome overfitting caused mainly by the lack of large labelled datasets while applying DNN, we propose to generate new realistic-like facial expressions using non-rigid registration techniques. The effectiveness of the suggested approach is demonstrated through conducting experiments on the BU-3DFE database. The quantitative evaluation and comparison with the recently developed state of the art show the competitiveness of the proposed 3D facial expression recognition approach.

Latent fingerprints are one of the most valuable and unique biometric attributes that are extensively used in forensic and law enforcement applications. Compared to rolled/plain fingerprint, latent fingerprint is of poor quality in term of friction ridge patterns, hence a more challenging for automatic fingerprint recognition systems. Considering the difficulties of dusting, lifting, and recovery of latent fingerprint, this type of fingerprints remain expensive to develop and collect. In this paper, we present a novel approach for synthetic latent fingerprint generation using Generative Adversarial Network (GAN). Our proposed framework, named mask to latent fingerprint (Mask2LFP), uses binary mask of distorted fingerprint-like shapes as input, and outputs a realistic latent fingerprint. This work focuses on the generation of synthetic latent fingerprints. The aim is to alleviate the scarcity issue of latent fingerprint data and serve the increasing need for developing, evaluating, and enhancing fingerprint-based identification systems, especially in forensic applications.

Infrared imaging has recently played an important role in a wide range of applications including video surveillance, robotics and night vision. However, infrared cameras often suffer from some limitations, essentially about low-contrast and blurred details. These problems contribute to the loss of observation of target objects in infrared images, which could limit the feasibility of different infrared imaging applications. In this paper, we mainly focus on the problem of pedestrian detection on thermal images. Particularly, we emphasis the need for enhancing the visual quality of images before performing the detection step. To address that, we propose a novel thermal enhancement architecture called TE-GAN based on Generative Adversarial Network, and composed of two modules contrast enhancement and denoising with a post-processing step for edge restoration in order to improve the overall image quality. The effectiveness of the proposed architecture is assessed by means of visual quality metrics and better results are obtained compared to the original thermal images and to the obtained results by other existing enhancement methods. These results have been conducted on a subset of KAIST dataset that we make available to encourage research in this direction11https://github.com/AmineMarnissi/TE-GAN, Using the same dataset, the impact of the proposed enhancement architecture has been demonstrated on the detection results by obtaining better performance with a significant margin using YOLOv3 detector.

Text line localization in document images is still considered an open research task. The state-of-the-art methods in this regard that are only based on the classical image analysis techniques mostly have unsatisfactory performances especially when the document images i) contain significant degradations and different noise types and scanning defects, and ii) have touching and/or multi-skewed text lines or overlapping words/characters and non-uniform inter-line space. Moreover, localizing text in ancient handwritten Arabic document images is even more complex due to the morphological particularities related to the Arabic script. Thus, in this paper, we propose a hybrid method combining a deep network with classical document image analysis techniques for text line localization in ancient handwritten Arabic document images. The proposed method is firstly based on using the U-Net architecture to extract the main area covering the text core. Then, a modified RLSA combined with topological structural analysis are applied to localize whole text lines (including the ascender and descender components). To analyze the performance of the proposed method, a set of experiments has been conducted on many recent public and private datasets, and a thorough experimental evaluation has been carried out.

Thanks to their interesting characteristics in terms of energetic performances and safety, Li-ion batteries Lithium Ferro-Phosphate (LFP) type are increasingly embedded in Electric Vehicles (EV). Hence a great interest to guarantee a high autonomy of the vehicle respecting the specific behavior of LFP batteries with accurate state of charge (SoC) monitoring. In this paper, we propose a full polynomial parameters-model of the battery first order model. The proposed model combined with the Extended Kalman Filter (EKF) is then used to estimate accurately the SoC of LFP battery. This monitoring method has been validated for Dynamic Discharge Pulse (DDP) profile, thus a high accuracy of SoC estimation is recorded, in fact an average error about 0.05% of SoC is obtained.

The attacks considered for keystroke dynamics study especially adaptive strategies have commonly treated impersonation attempts known as zero-effort attacks. These attacks are generally the acquisition of other users of the same database while typing the same password without intending to impersonate the genuine user account. To deal with more realistic scenarios, we are interested in this paper to study the robustness of an adaptive strategy against four types of imposter attacks: zero-effort, spoof, playback and synthetic applied to the WEBGREYC database. Experimental results show that 1) playback and synthetic attacks are the most dangerous and increase the EER rates compared to the other attacks; 2) we also find that the impact of these attacks is more pronounced when the percentages of imposter samples are greater than those of genuine ones; 3) the spoof attacks achieve alarmingly higher FMR, FNMR, and EER rates compared to zero-effort impostor attacks; 4) FMR, FNMR, and EER are higher when the percentage of attacks increases; 5) the attacks belonging to the same user are more dangerous than those of different users in particular when the percentage of the attacks increases. In light of our results, we point out that the traditional attacks considered in research on keystroke-based authentication must evolve according to the evolution of the attacks of nowadays password-based applications.

On most document image transcription, indexing and retrieval systems, text line segmentation remains one of the most important preliminary task. Hence, the research community working in document image analysis is particularly interested in providing reliable text line segmentation methods. Recently, an increasing interest in using deep learning-based methods has been noted for solving various sub-fields and tasks related to the issues surrounding document image analysis. Thanks to the computer hardware and software evolution, several methods based on using deep architectures continue to outperform the pattern recognition issues and particularly those related to historical document image analysis. Thus, in this paper we present a novel deep learning-based method for text line segmentation of historical documents. The proposed method is based on using an adaptive U-Net architecture. Qualitative and numerical experiments are given using a large number of historical document images collected from the Tunisian national archives and different recent benchmarking datasets provided in the context of ICDAR and ICFHR competitions. Moreover, the results achieved are compared with those obtained using the state-of-the-art methods.

This paper serves as a survey and empirical evaluation of the state-of-the-art in activity recognition methods using still RGB images and/or videos. Understanding human activities from videos or still images is a challenging task in computer vision domain. Identifying the action or activity being accomplished automatically and then recognizing it represents the prime goal of an intelligent video system. Human Activity Recognition arises in various application domains varying from human computer interfaces, health care monitoring to surveillance and security. Despite the ongoing efforts in the domain, these tasks remained unsolved in unconstrained environments and face many challenges such as occlusions, variations in clothing and background clutter. Recently, numerous deep learning algorithms have been proposed to solve traditional artificial intelligence problems. They have shown great advances, in particular for pose estimation task since they can extract appropriate features while jointly performing discrimination. In this paper, we provide a detailed review of recent and state-of-the-art research advances in the field of human activity recognition. We propose a categorization of human activity methodologies and discuss their advantages and limitations. In particular, we divide feature representation methods into global, local and body modeling. Then, human activity classification approaches are arranged into three categories, which reflect how they model human activities: template-based, generative and discriminative. Moreover, we provide a comprehensive analysis of pose-based human activity recognition where both conventional and deep learning-based human pose estimation approaches are reported. Finally, we discuss the open-challenges in this field and endeavor to provide possible solutions.

Cyber-attacks have spread all over the world to steal information such as trade secrets, intellectual property and banking data. Facing the danger of the insecurity of saved data (personal, professional, official, etc.), keystroke dynamics was proposed as an interesting, non-intrusive, inexpensive, permanent and weakly constrained solution for users. Based on the typing rhythm of users, it improves logical access security. Nevertheless, it was demonstrated that such an authentication mechanism would need a larger number of samples to enroll the typing characteristics of users. Moreover, these registered characteristics generally undergo aging effects after a time span. Different solutions have been suggested to remedy these variability problems, including template adaptation. In this paper, we propose a double serial adaptation strategy that considers a single-capture-based enrollment process. When using the authentication system, the template of users and the decision/adaptation thresholds are updated. Experimental results on three public keystroke dynamics datasets show the benefits of the proposed method.

Intelligent traffic systems for traffic surveillance and monitoring have become a topic of great interest to some cities in the world. Generally, the existing traffic surveillance systems are made up of costly equipment with complicated operational procedures and have difficulties with congestion, occlusion, and lighting night/day and day/night transitions. In this paper, we propose an embedded system for traffic surveillance that can be utilized under these challenging conditions. This system analyses traffic and particularly focuses on the problem of detecting and categorizing traffic objects in several traffic scenarios. Moreover, it contains a robust detector produced by an original specialization framework. The proposed specialization framework utilizes a generic deep detector so as to improve the detection accuracy in a specific traffic scenario. The experiments demonstrate that the proposed specialization framework presents encouraging results for multi-traffic object detection and outperforms the state-of-the-art specialization frameworks on several public traffic datasets.

Regarding the fact that individuals have different interactions with biometric authentication systems, several techniques have been developed in the literature to model different users categories. Doddington Zoo is a concept of categorizing users behaviors into animal groups to reflect their characteristics with respect to biometric systems. This concept was developed for different biometric modalities including keystroke dynamics. The present study extends this biometric classification, by proposing a novel adaptive strategy based on the Doddinghton Zoo, for the recognition of the user's keystroke dynamics. The obtained results demonstrate competitive performances on significant keystroke dynamics datasets.

Securing personal, professional and even official data is a very critical issue nowadays, giving that these informations are safeguarded in different devices (mobile, computer) and various accounts (social networks, e-mails). To protect them from unauthorized access, users generally are asked to use passwords. But using only this authentication solution is no longer efficient against hacker attacks. Keystroke dynamics is a biometric promising modality that guarantees the recognition of the user's characteristics; his typing manner on the keyboard. Regarding that the typing rhythm of the user changes over time, adaptive biometric strategies help to take into consideration these variations during the authentication system. In this paper we classify users into multiple categories according to Doddington Zoo classification. Afterwards, we apply an adaptive strategy specific to each category of users. The achieved experiments demonstrate that an update strategy specific to the user class significantly improves the obtained performances.

This study presents a novel approach for Arabic video text recognition based on recurrent neural networks. In fact, embedded texts in videos represent a rich source of information for indexing and automatically annotating multimedia documents. However, video text recognition is a non‐trivial task due to many challenges like the variability of text patterns and the complexity of backgrounds. In the case of Arabic, the presence of diacritic marks, the cursive nature of the script and the non‐uniform intra/inter word distances, may introduce many additional challenges. The proposed system presents a segmentation‐free method that relies specifically on a multi‐dimensional long short‐term memory coupled with a connectionist temporal classification layer. It is shown that using an efficient pre‐processing step and a compact representation of Arabic character models brings robust performance and yields a low‐error rate than other recently published methods. The authors’ system is trained and evaluated using the public AcTiV‐R dataset under different evaluation protocols. The obtained results are very interesting. They also outperform current state‐of‐the‐art approaches on the public dataset ALIF in terms of recognition rates at both character and line levels.

In mammograms, the breast skin line often appears ambiguous and poorly defined. This is mainly due to the breast organ compression during the image acquisition process along with the inherent low density of the tissue in that area. The accurate delimitation of the breast region becomes a challenging task to conventional segmentation techniques. In this study, the authors propose a new segmentation approach allowing to overcome this challenge. This approach is based on the application of two complementary segmentation techniques exploring each, respectively, the grey‐scale intensities and the local‐homogeneity domains. The knowledge resulting from each segmentation technique is considered as a knowledge source and is modelled using the belief functions formalism. The two considered knowledge sources are then fused using an iterative process. The obtained results show the efficiency of the proposed evidential approach especially in terms of ambiguity removal and decision quality improvement for accurate breast border delimitation (which is often under‐segmented and assimilated to the background by most of the existing segmentation techniques).

3D virtual worlds have technologically emerged as platforms able to support millions of users to interact with each other, sharing information on business, education, social science, news and cutting-edge technologies. Despite their maturity, they are subject to cybercrimes due to the lack of an efficient access control mechanism. Technically, it is l essential to install a biometric access control solution to secure these virtual spaces. Actually, virtual world platforms are only equipped with classical password-based authentication mechanism. The implementation of new a logical access control procedure in publicly accessible virtual worlds is closely dependent on the characteristics of the virtual world platform. In this study, we put forward a methodology to design a virtual world platform for the implementation of a biometrical access control mechanisms. For virtual world building, we use as software a set of 3D tools and the Unreal Engine, which is a complete suite of 3D game and space development tools. To implement the access control application, we develop a centralized biometric authentication module based on fingerprints.

In this paper we propose a novel method based on Recurrent Neural Networks (RNN) for text recognition in Arabic news video frames. In fact, embedded texts in videos represent a rich source of information for indexing and automatic processing of multimedia documents. However, text recognition in video is not trivial due to many challenges like background complexity (e.g., presence of text-like objects), unknown text size/font with various colours and degraded text quality. The proposed system presents a segmentation-free recognition technique (i.e. no prior required segmentation of words into characters) using a RNN architecture. This technique relies specifically on a Multi-Dimensional Long Short Term Memory (MDLSTM) with a Connectionist Temporal Classification (CTC) output layer. Our system has been evaluated on the public AcTiV-R dataset. The obtained results are very promising.

This paper describes the multi-resolution Arabic Text detection and recognition in Video Competition-AcTiVComp held in the context of the 14thInternational Conference on Document Analysis and Recognition (ICDAR' 2017), during November 9-15, 2017, in Kyoto, Japan. The main objective of this competition is to evaluate the performance of participants' algorithms for automatically detecting and recognizing Arabic texts in video frames using the freely available Arabic-Text-in-Video (AcTiV) dataset. A first edition was held in the framework of the 23rdInternational Conference on Pattern Recognition (ICPR'2016). Three groups with five systems are participating to the second edition of AcTiVComp. These systems are tested in a blind manner on a closed-subset of the AcTiV database, which is unknown to all participants. In addition to the experimental setup and observed results, we also provide a short description of the participating groups and their systems.

This study addresses the problem of efficiently combining the joint, RGB and depth modalities of the Kinect sensor in order to recognise human actions. For this purpose, a multi‐layered fusion scheme concatenates different specific features, builds specialised local and global SVM models and then iteratively fuses their different scores. The authors essentially contribute in two levels: (i) they combine the performance of local descriptors with the strength of global bags‐of‐visual‐words representations. They are able then to generate improved local decisions that allow noisy frames handling. (ii) They also study the performance of multiple fusion schemes guided by different features concatenations, Fisher vectors representations concatenation and later iterative scores fusion. To prove the efficiency of their approach, they have evaluated their experiments on two challenging public datasets: CAD‐60 and CGC‐2014. Competitive results are obtained for both benchmarks.

Fuzzy median graph is an important new concept that can represent a set of fuzzy graphs by a representative fuzzy graph prototype. However, the computation of a fuzzy median graph remains a computationally expensive task. In this paper, we propose a new approximate algorithm for the computation of the Fuzzy Generalized Median Graph (FGMG) based on Fuzzy Attributed Relational Graph (FARG) embedding in a suitable vector space in order to capture the maximum information in graphs and to improve the accuracy and speed of document image retrieval processing. In this study, we focus on the application of FGMGs to the Content-based Document Retrieval (CBDR) problem. Experiments on real and synthetic databases containing a large number of FARGs with large sizes show that a CBDR using the FGMG as a dataset representative yields better results than an exhaustive and sequential retrieval in terms of gains in accuracy and time processing.

Text detection in videos is a challenging problem due to variety of text specificities, presence of complex background and anti-aliasing/compression artifacts. In this paper, we present an approach for horizontally aligned artificial text detection in Arabic news video. The novelty of this method revolves around the combination of two techniques: an adapted version of the Stroke Width Transform (SWT) algorithm and a convolutional auto-encoder (CAE). First, the SWT extracts text candidates' components. They are then filtered and grouped using geometric constraints and Stroke Width information. Second, the CAE is used as an unsupervised feature learning method to discriminate the obtained textline candidates as text or non-text. We assess the proposed approach on the public Arabic-Text-in-Video database (AcTiV-DB) using different evaluation protocols including data from several TV channels. Experiments indicate that the use of learned features significantly improves the text detection results.

This paper describes the AcTiVComp: detection and recognition of Arabic Text in Video competition in conjunction with the 23rd International Conference on Pattern Recognition (ICPR). The main objective of this competition is to evaluate the performance of participants' algorithms to automatically locate and/or recognize overlay text lines in Arabic video frames using the freely available AcTiV dataset. In this first edition of AcTiVComp, four groups with five systems are participating to the competition. In the detection challenge, the systems are compared based on the standard assessment metrics (i.e. recall, precision and F-score). The recognition results evaluation is based on the recognition rates at the character, word and line levels. The systems were tested in a blind manner on the closed-test set of the AcTiV dataset which is unknown to all participants. In addition to the test results, we also provide a short description of the participating groups and their systems.

In this paper we propose a new approach for watermarking biometric fingerprint images using Gabor direct linear discriminant analysis face features. Our goal is to incorporate a watermark in the best embedding domain that preserves minutiae, which are the most relevant proven features of a fingerprint. We conducted a comprehensive study based on the influence of the watermark embedding domain choice on the performance of the minutiae-based identity and of the robustness and imperceptibility of the watermarking approach. Three embedding domains were tested: spatial, frequency and multiresolution. The various tests were performed on two biometric databases, multimodal and chimerical. The best results were recorded in the multiresolution domain in terms of preserving the minutiae number and positions. Moreover, this embedding domain led to the best verification performances and to a good compromise between robustness and imperceptibility.

The evolution of digital technology and the desire to maintain an easy access to scanned documents implies the interest of Document Image Retrieval (DIR). In this paper, we propose a fuzzy graph-based document retrieval approach for determining similarity between document images for content-based document retrieval. To model the structure of document images, we have opted to use Attributed Relational Graphs (ARGs). For each document region (text, graphic, etc) we associate a node which is characterized by a set of fuzzy membership degrees reflecting low-level properties (texture, shape, color, etc). We use this fuzzy description in order to guarantee more robustness against the eventual segmentation errors which may be occurred after the segmentation of document regions. ARGs edges represent spatial relationships between regions which have common boundaries. Finally, we have developed a tree-search based optimal matching algorithm, which allows the search for document according to its structure. The database used for experiments is composed of segmented Coran document images from the National Library of Tunisia. Different weights are assigned for regions and edges according to their relative importance. The results obtained demonstrate the effectiveness of fuzzy graph-based document retrieval. Such an approach is very useful for several applications in many fields.

Recently, promising results have been reported on video text detection and recognition. Most of the proposed methods are tested on private datasets with non-uniform evaluation metrics. We report here on the development of a publicly accessible annotated video dataset designed to assess the performance of different artificial Arabic text detection, tracking and recognition systems. The dataset includes 80 videos (more than 850,000 frames) collected from 4 different Arabic news channels. An attempt was made to ensure maximum diversities of the textual content in terms of size, position and background. This data is accompanied by detailed annotations for each textbox. We also present a region-based text detection approach in addition to a set of evaluation protocols on which the performance of different systems can be measured.

We present in this paper a state of the latest advances in the field of offline handwritten signature verification. We describe the main approaches that have been proposed in the recent decades. Besides, we introduce the database of static signatures published in the literature as well as international competitions organized in the domain. Also, we present our contribution in the field.

Virtual worlds have become a target to criminal activities. With the increasing interest in these spaces, their security is becoming a challenge. The recently published works have proposed methodologies for the security of the virtual world based on avatar face authentication. However, the developed recognition systems have been validated on the avatar image datasets which belong to the authors themselves. In order to objectively judge these methods, a great availability of standard datasets is required. Taking into account the importance of the considered stakes, the development of a common database to test the different suggested techniques is becoming an essential step. In this paper we present collected datasets from the Second Life virtual world which contains 3D meshes of avatars representing the real objects in the virtual worlds. Our work consists in conceiving a standard sub-base that can be used to develop new security methodologies dedicated to avatar recognition or authentication in the 3D dimension.

In this paper, we put forward a new ground truth database of 500 real ancient document images suffering from different degrees of an ink bleed-through degradation. Each image is associated with a semi-automatic generated ground-truth mask. The conception and creation of this database is described. In addition we compare three single-side image restoration approaches in two ways: visually and quantitatively using the generated ground truth masks.

In this paper, we propose a new approach based on watermarking for fusing biometric modalities. The main idea of the proposed approach is to use the watermark both for security purpose and as an additional information related to the person, therefore increasing the personal data used for verification. The host image of the face is watermarked in the multi-resolution space by the palmprint using a watermarking technique based on wavelet packet decomposition. For the verification stage, the characterization of the watermarked face is provided by Gabor filters while classification is performed by SVMs. The experimental results show that this technique ensures a significant performance improvement in both identity verification and biometric security over the use of a single system.

Hidden Markov Models (HMMs) are now widely used for off-line Arabic handwriting recognition. Actually, classical HMMs are one-dimensional models, that is why to process an Arabic word image we have developed a discrete Dynamic Bayesian Network (DBN). The DBNs are an extension and a generalization of the classical HMMs, which can model the interaction between several observations and state sequences. In our study, we have represented words by factorizing two streams in different manners, where the interaction is achieved through the causal influence between observable variables in the first model and state variables in the second one. The aim of this is to consider the two flows of information together: The observations on the columns (as well as lines) are obtained by scanning the image horizontally (and also vertically) by a uniform sliding window. We have compared the two models on the recognition of off-line Arabic handwritten words. The experiments show that the first model is better and more adapted to our task than the second one.

We present in this paper a novel solution for temporal segmentation of human gestures that takes advantage of the skeletal-joints streams offered by the Kinect sensor. Our contribution consist in introducing an improved skeletal representation and its usage in a multilayer motion delimitation that distinguishes the non-vocabulary actions. The evaluation of the solution is presented on a subset of the Chalearn Gesture Challenge (CGC) 2014 dataset. The obtained temporal segmentation is better than the CGC baseline methods and has proved to be important for the task of human-action recognition.

The ancient documents represent one of the author's history pillars that preserve the historical events for the next generation. Many digitisation projects have been launched to preserve and to diffuse them to the public. However, ensuring security is a challenge since digital documents are susceptible to be hacked. Thus, watermarking the images of the documents seems to be a promising solution mainly for protect the copyright. Many watermarking algorithms have been proposed particularly in the medical field. However, no solution has been recommended for the images of ancient documents. In this study, the authors present a watermarking approach devoted to store images. Their algorithm is based on a wavelet packet decomposition, an intelligent choice of the best base carrier points by way of the singular value decomposition. In their approach, the insertion base and the carrier points of the signature are dynamic, varying from one image to another. For a better preservation of the watermark, they exploited a convolutional encoder to encode data. The results recorded in a set of images of ancient documents taken from the National Library of Tunisia and from the National Archives of Tunis have shown promising results.

In classification tasks, class-modular strategy has been widely used. It has outperformed classical strategy for pattern classification task in many applications. However, in some modular architecture, such as one against all in support vector machines classifier, the training dataset for one class risks to heavily outnumber the other classes. In this challenging situation, the trained classifier will accurately classify the majority class; nevertheless, it marginalizes the minority class. As a result, True Negatives rate (TNr) will be very high while the True Positives rate (TPr) will be low. The main goal of this work is to improve TPr without much sacrifice in TNr. In this paper, we propose oversampling the minority class using polynomial fitting functions. Four new approaches were proposed: star topology, bus topology, polynomial curve topology and mesh topology. Star and mesh topologies approach had led to the best performances.

In this paper, we propose a method for features extraction from multifont Arabic characters images based on the Contourlet Transform, which has been recently introduced. In our previous works, we noticed that Wavelet transforms are not capable of reconstructing curved images perfectly; the Contourlet Transform offers a solution to remedy to this insufficiency. It allows a multiresolution and directional decomposition of a signal using a combination of Laplacian Pyramid (LP) and a Directional Filter Bank (DFB). The Contourlet Transform has good approximation properties for smooth 2D functions and finds a direct discrete-space construction, and is therefore computationally efficient. Experimental tests have been carried out on a set of 175.000 samples of characters corresponding to 9 different Arabic fonts. Some promising experimental results are reported.

In this paper, we present an approach for writer identification using off-line Arabic handwriting. The proposed method explores the handwriting texture analysis by 2D discrete wavelet transforms using lifting scheme. A comparative evaluation between textural features extracted by 9 different wavelet transform functions was done. A modular multilayer perceptron classifier was used. Experiments have shown that writer identification accuracies reach best performance levels with an average rate of 95.68%. Experiments have been carried out using a database of 180 text samples. The chosen text was made to guarantee the involvement of the various internal shapes and letter locations within an Arabic subword.

The recognition of the cursive writing requires generally a segmentation procedure. This operation constitutes one of the major difficulties of all OCR systems. The research of an ideal segmentation process of a word into letters is Utopian. In this paper we propose two segmentation-by-recognition techniques for cursive printed Arabic recognition based on Hough transform.

A novel approach for the Arabic handwriting recognition is presented. The use of a planar hidden Markov model (PHMM) has permitted to split the Arabic script into five homogeneous horizontal regions. Each region was described by a 1D-HMM. This modeling is based on different levels of segmentation: horizontal, natural and vertical. Both holistic and analytical approaches have been tested for the description of the median band of the Arabic writing. We show finally that a hybrid approach conducted to the improvement of the whole system performances.

The Generalized Hough Transform is a technique used to detect arbitrary objects in a given image. This technique is known for its capacity of absorption of distortions as well as noises. In the present paper, we describe an approach showing the efficiency of the use of the Generalized Hough Transform to recognize Arabic printed characters in their different shapes.

In this paper, we show how planar hidden Markov models (PHMM) can offer great potential to solve difficult Arabic character recognition problems, especially its cursivness. A convenient architecture is defined for printed Arabic sub-words. It yields an easy solution to implement the modeling of the different morphological variations of the Arabic writing, i.e., vertical and variable horizontal linkages. A more flexible architecture, developed for Arabic handwritten words, is under test. The structure proposed presents the aptitude to absorb the variability of the manuscript. Indeed, the experiments have shown promising results and directions for further improvements. In the present paper, we describe both retained architectures, showing the applicability of the PHMM to the Arabic complexities. This is owed precisely to the definition of the PHMM, which permits to follow efficiently the natural variations in bands of the Arabic script.

In this paper, we present an approach for connected Arabic printed text recognition using statistical models based on planar hidden Markov models (PHMM), without prior segmentation. The performance is enhanced by the use of robust features and an efficient superstate duration distribution. The approach has been tested on a vocabulary of 11 kinds of pieces of arabic word (PAW) of three characters each. The experiments have shown promising results and directions for further improvements. The recognition accuracy has proved to be of 100% even with poor and degraded texts.

In this paper we present the segmentation module of a complete system for the recognition of printed and handwritten Arabic documents. The system which is under development, includes several modules especially for characterising text fonts. It is based on the use of the Hidden Markov Models together with decision trees and a dictionary correction system. After a brief overview of the recent work in the field, a new methodology for segmenting off-line Arabic printed characters is presented. This approach is simple and very efficient. The paper describes the choice of the different primitives that can be extracted from the image of the character and the use of the modulated histogram as well as the number of black segments in a line of pixels. The paper identifies various sources of errors and factors that make the task perfect segmentation, difficult. The present algorithm has been tested with most of the print fonts and is currently being tested for handwritten characters. The results obtained are very promising.