Hand Gesture Recognition (HGR) has gained significant attention as a natural and intuitive means of human-computer interaction, driven by advances in machine learning, sensor technologies, and computational power. Among these, the Leap Motion Controller (LMC) stands out for its ability to capture precise hand motion with high spatial accuracy and multiple modalities (skeletal and depth). In this work, we demonstrate that representing these complex multimodal outputs as graph structures is not only appropriate but also a significant innovation that fully utilizes sophisticated neural network architectures. To improve accuracy and generalization in HGR tasks, we present a hierarchical transformer gated graph neural network that has been improved with an intermediate fusion technique. Our architecture has been designed to overcome several limitations of current graph-based methods, including inadequate multi-modal feature integration, insufficient temporal modeling, and limited generalization across users. Using two benchmark datasets, 2MLMD and MMHGD, we examine our approach and find consistent and substantial enhancements in performance compared to the most advanced baselines. Superior accuracy, robustness, and generalizability are demonstrated by the results, confirming the efficacy of our architecture. Ablation experiments demonstrate the significance of each component in improving recognition performance.

Ultrasound imaging is widely used in clinical practice due to its non-invasive, real-time, and cost-effective nature. However, speckle noise often degrades image quality, obscuring fine anatomical structures and reducing diagnostic confidence. Existing denoising methods struggle to remove noise effectively while preserving critical details, limiting their clinical utility. Although recent deep learning architectures excel at capturing both local details and global structure, they remain inherently limited in handling speckle noise, as its physical characteristics are not explicitly incorporated. To address this limitation, a Physics-Regularized Self-Supervised Denoising U-Net (PR-SSD-Net) is introduced to reinforce the U-Net's capability for high-quality image restoration. The physics-based constraint guides the network to produce residual noise patterns that align with expected statistical behavior, enhancing image clarity and preserving critical structures. Comprehensive evaluations were conducted on six diverse ultrasound datasets. Significant improvements in Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM) were observed, accompanied by reduced variability, as reflected in their standard deviations (SD). An ablation study confirmed the pivotal role of physics-guided regularization, and expert assessments demonstrated high inter-rater agreement (Fleiss), supporting the clinical relevance of the approach. These results highlight the proposed PR-SSD-Net approach as a robust, physically grounded solution for speckle noise reduction, enhancing both the reliability and clinical utility of ultrasound imaging.

Person Re-Identification (ReID) remains a pivotal challenge in intelligent surveillance, constantly tested by dramatic intra-class variations, pose changes, lighting shifts, occlusions and distracting backgrounds. This paper proposes a self-attention based hybrid framework (VSPeR) designed to improve data diversity and visual quality for person re-identification in real-world scenarios. At the heart of VSPeR lies a powerful synergy: the robust, identity preserving feature extraction of Variational Autoencoders (VAEs) meets the precise, context-aware focus of Self-Attention Generative Adversarial Networks (GANs). This hybrid architecture captures rich, semantically meaningful latent representations while zeroing in on the most discriminative visual cues necessary for accurate cross-view matching. The outcome is high-fidelity, identity-consistent synthetic images that significantly amplify the training dataset enabling smarter, more effective learning. Put to the test across four leading ReID benchmarks, VSPeR achieves competitive performance compared to existing approaches in terms of mean Average Precision (mAP) and Rank-1 accuracy.

Person re-identification (ReID) is a core task in computer vision, requiring consistent identification of individuals across cameras despite variations in appearance, pose, lighting, and occlusion. Although deep learning methods have significantly improved performance, they remain sensitive to degraded image quality, dependence on large annotated datasets, and limited generalization under uncontrolled conditions. In this paper, we propose PId-SAGAN, a generative self-attention framework with identity preservation for person ReID. We introduce Final-Layer Identity Attention (FLIA), a lightweight self-attention mechanism applied exclusively at the generator’s final layer. This design enables the model to capture long-range dependencies, refine global structures, suppress background noise, and preserve fine-grained identity cues through an identity-preserving loss, while maintaining computational efficiency. The proposed framework integrates identity-aware image generation with CNN-based ReID learning using both real and synthetic data, improving robustness under challenging scenarios such as occlusion, illumination changes, and pose variations. Experimental results on multiple public benchmarks demonstrate that PId-SAGAN achieves competitive performance compared to recent generative and attention-based approaches in uncontrolled real-world person ReID.

Abstract Person re-identification (ReID) poses significant challenges in practical surveillance scenarios, where identity-discriminative information is often degraded by occlusion, low resolution, motion blur, and varying illumination. While diffusion-based models have recently shown strong potential in image generation, their use in ReID has been largely limited to data augmentation, leaving their potential for directly restoring identity-relevant information from degraded probe images largely unexplored. In this paper, we introduce PerDiff , a novel diffusion-driven framework for identity-preserving reconstruction in person ReID. Unlike conventional generative approaches, PerDiff reformulates diffusion as a task-driven, identity-aware inverse process, where corrupted person images are progressively denoised to recover identity-discriminative visual cues prior to feature extraction. By tightly coupling diffusion-based reconstruction with identity recognition, PerDiff produces identity-consistent representations that are explicitly optimized for ReID matching rather than visual realism alone. Extensive evaluations on four benchmark datasets, including Market-1501, CUHK03, DukeMTMC-reID, and IUST_PersonReID, demonstrate that PerDiff consistently improves ReID performance in both mean Average Precision and Rank-1 scores, particularly under severe degradation conditions. These results establish diffusion-based identity reconstruction as a powerful new paradigm for robust person ReID beyond traditional data augmentation strategies.

Speckle noise in ultrasound images compromises image quality and hinders diagnostic accuracy. Traditional ultrasound denoising methods often struggle to preserve anatomical details while effectively reducing noise, especially under high-noise conditions. In this study, we propose an innovative approach that integrates a lightweight channel attention mechanism (LCAM) within a convolutional variational autoencoder (CVAE) framework to enhance ultrasound image denoising. The proposed approach efficiently reduces speckle noise while maintaining essential anatomical features. Comprehensive evaluations across six diverse ultrasound datasets demonstrate that the LCAM-CVAE outperforms conventional denoising techniques in both subjective image quality and objective performance metrics, including peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), standard deviation in PSNR (SD-PSNR), standard deviation in SSIM (SD-SSIM), PSNR statistical relevance tests, and computational efficiency (CE). The LCAM-CVAE approach demonstrates exceptional performance, particularly under high-noise conditions, ensuring the preservation of key anatomical structures for accurate diagnosis. These results highlight the LCAM-CVAE approach as a robust and promising solution for ultrasound image denoising, with significant clinical potential to improve diagnostic quality in noisy environments.

With the rapid technological progress, drones, or unmanned aerial vehicles (UAVs), have emerged among the most important artificial intelligence (AI)-powered systems. With their aerial perspective, mobility, and cost-effectiveness, they became crucial for advancing AI-driven visual perception in various sectors. However, implementing generic object detection algorithms on these resource-limited devices remains a complex challenge. Towards efficient and more UAV-adapted systems, this paper introduces the Faster Real-time Detector based on You Only Look Once (FRD-YOLO). FRD-YOLO presents different optimizations on the functional and architectural perspectives. To adapt the model to the vision context of UAVs, our key enhancements include the addition of a new layer for detecting tiny objects and the removal of the detection layer for large objects to emphasize the small and tiny targets. We also introduce a structure-aware integration of C3Ghost blocks, inspired by Ghost Convolutions and Cross Stage Partial Network-based layers, to reduce computational cost, and integrate the Convolutional Block Attention Module to enhance the recall. The FRD-YOLO model demonstrates superior detection performance with reduced size and computations across its five scaled versions. Notably, the evaluation on the challenging VisDroneDet2021 dataset reveals that the FRD-YOLO-x achieves 11.23% higher mean Average Precision (mAP50) than the baseline model, with 25.44% less computational cost, reaching up to 44 Frames Per Second (FPS). Additionally, the FRD-YOLO model showcases reliable embedded inference on the Jetson TX2, with FRD-YOLO-n achieving 25.18 FPS and FRD-YOLO-x reducing model size by 58.1%, confirming the architecture’s strength for UAV deployment.

In ultrasound imaging, speckle noise is acknowledged as a major challenge, affecting visual clarity and compromising diagnostic accuracy. This degradation can lead to erroneous clinical evaluations, highlighting the urgent need to enhance denoising techniques. Various solutions have been proposed to tackle this issue. Convolutional Neural Network-based Denoising AutoEncoders (CNN-DAE) aim to mitigate this issue. However, this strategy suffers from potential loss of intricate details, excessive smoothing of images, and limited ability to adapt to different noise characteristics. In this research, we propose a CNN based Edge Detection Denoising AutoEncoders (CNN-EDDA). The proposed approach focuses on speckle denoising and enhancing the contours of anatomical structures, providing sharper and more precise definitions. This clarity enables the algorithm to effectively denoise the image by identifying regions with abrupt variations, allowing for the distinction of important details from speckle noise. By clarifying the edges, we preserve essential image details while reducing noise. This synergy between filtering and edge detection is crucial for ensuring reliable, high-quality results, facilitating more effective clinical interpretation. Our proposed CNN-EDDA method was evaluated according to subjective visual analyses as well as objective measurements, demonstrating better results than current state-of-the-art methods, with optimal performance metrics.

Vehicle Re-Identification involves identifying and matching a target vehicle with images captured from different views in a multi-camera network. This topic holds significant importance in various applications including intelligent transportation systems, video surveillance and smart city. However, Vehicle Re-Identification faces significant challenges in dynamic environments due to viewpoint variations, inter-vehicle appearance similarity, intra-class variability, illumination variation, occlusion and background clutter. To address these limitations, we propose HA-VReID, An Effective Hard Attention Model with Deep Learning for Vehicle Re-Identification that combines Hard attention mechanism for background removal and vehicle shape focus with EfficientNet-powered feature extraction for robust vehicle representation. Extensive experiments on the VeRi-776 and VRAI benchmarks demonstrate that our approach outperforms state-of-the-art methods in vehicle Re-Identification tasks.

The 3D hand skeletal data has received considerable amount of attention due to its potential uses case for Hand Gesture Recognition (HGR) systems. Additionaly, Graph Neural Network (GNN) have been widely employed for skeletal-based HGR. Yet still, traditional models frequently suffer from inefficient feature representation and generalizability. Thus, to overcome these limitations, we introduce Spatial Graph Isomorphism Neural Networks (S-GINs), which use GIN layers to improve the feature aggregation. First, we create a graph representing skeletal-based recordings. Following that, a spatial network convolution module learns the inherent topology of hand gestures from neighbor nodes and updates it using a multilayer perceptron. This method promotes classification accuracy over other spatial based graph models including graph attention network, graph sampeling and aggregating, and graph convolutional networks. We validate S-GINs employing three benchmark datasets: MMHGD, 2MLMD, and Multi-view Leap2 , both horizontal and vertical sub-datasets. Experimental findings show that our model outperforms state-of-the-art graph based methods, by elevating the accuracies to 72%, 81%, 85%, and 83% respectively .

Traditional input methods in Virtual Reality (VR) cause physical discomfort. However, touchless technologies using Hand Gesture Recognition (HGR) do not provide sufficient reliability or convenience for smooth operation. Only four HGR datasets are now publicly available for VR, indicating shortcomings in their handling of dynamic and bimanual gestures. Furthermore, the majority of them are based on static gestures that are ineffective for VR systems. To address these limitations, we introduce the Virtual Reality Gestures Dataset (VR-GD) as a new, comprehensive, and challenging dataset. Our novel dataset contains 27,600 samples from 30 participants and records 46 gestures, including 41 dynamic and 30 bimanual gestures, through a head-mounted Leap Motion Controller (LMC) providing detailed skeletal and depth information. The gestures are applied throughout VR environments, including VR automotive simulation, system control, and gaming. Among the models evaluated for HGR were Inception V3 and the Spatiotemporal Graph Neural Network, with Inception V3 demonstrating the best results by achieving 86.2% accuracy and an 86.0% F1-score, highlighting its effectiveness in capturing complex gesture patterns.

Transfer learning is considered an effective technique that enhances model performance by using knowledge acquired from a source dataset to tackle a similar task on a target dataset. This technique is particularly valuable in fields where labeled data is limited, such as thermal comfort prediction. In the Tunisian context, the lack of specific thermal comfort data for office spaces occupied by multiple people represents a major challenge to optimizing workplace environments. To address this, we introduce a transfer learning-based approach using ASHRAE RP-884 data from countries with similar climatic conditions. In fact, these data were purified based on climate conditions and selected for office-type buildings. Three transfer learning methods were evaluated using three models: a Deep GRU-BiGRU model, a Deep GRU model, and a BiGRU model. Our results present that the transfer learning approach based on the Deep GRU-BiGRU model achieves the highest accuracy, reaching 66.15% in thermal comfort prediction, outperforming the other methods.

Emotion recognition from text is a key task in natural language processing (NLP), with applications ranging from interactive chat systems and mental health detection to consumer feedback analysis. While various machine learning and deep learning techniques have been explored, the success of Large Language Models (LLMs) presents new opportunities revolutionizing how we interact with text. In this study, we harness LLaMA-3-8B for this task by applying LoRA (Low-Rank Adaptation), a highly effective fine-tuning technique that reduces computational and memory costs while maintaining high performance. Experiments on the ISEAR, Emotion for NLP and SemEval 2019 datasets demonstrate significant improvements in accuracy and efficiency over traditional approaches. This work highlights the potential of Llama-3-8B, combined with LoRA, to excel in complex language understanding tasks, particularly in emotion recognition.

Convolutional neural networks attract the highest research focus in the developing field of Hand Gesture Recognition (HGR). Nevertheless, these approaches presented a challenging task in adapting to time-series data. In skeleton-based HGR, extracting spatial-temporal information remains a challenge. In recent times, recurrent neural networks have exhibited exceptional performance in detecting desired hand gestures by processing of varied length time-series data. Although they outperform traditional methods when huge training data is accessible, their effectiveness significantly diminishes when data availability is constrained. In this study, we introduce an unsupervised data augmentation network known as the Spatial-Temporal Generative Network (STGN), which reconstructs both the spatial and temporal information of the input sequences by leveraging a Deep Long Short-Term Memory Auto-Encoder (DLSTM-AE) network. Consequently, the DLSTM-AE combined with different Long Short-Term Memory (LSTM) network variations, forming an integrated network that can be trained end-to-end for HGR. Through experimentation conducted on the LeapGestureDB dataset (Leap Motion-based Gesture Dataset) and RIT dataset (Rochester Institute of Technology Hand Gesture Dataset), we prove that data reconstruction using STGN had a prominent effect on improving the accuracy of recognizing time-series based hand gestures. For all experiments, the best recognition results are achieved in the augmented dataset. Accuracies were improved on all tested LSTM networks from 2 to 10%. For reproducible research, the code is available at: https://github.com/AMEURsafa/STGN.

Integrating deep learning-based object detection into Unmanned Aerial Vehicles (UAVs) has emerged as a cutting-edge solution in modern technology for real-time perception and surveillance tasks. However, deploying these models on resource-constrained edge devices presents substantial challenges. While high-performance platforms support models with large parameter counts and high computational demands, which offer superior accuracy, such complexity significantly compromises real-time inference on low-power embedded systems. To achieve a balance between precision and real-time performance, we propose an optimized drone target detection model based on YOLOv10. First, we incorporate the C3k2 module, which applies spatial and channel feature fusion through parallel lightweight 3×3 convolutions to reduce the computational load of feature extraction. Secondly, we integrate the A2C2f module, which applies an area-based attention mechanism to improve the accuracy. Experimental findings on the public VisDrone dataset indicate that the optimized YOLOv10 model achieves a reduction in computational load by 10.61%, improves the inference speed by 3.33%, and maintains mAP@50 at 37%. Moreover, when evaluated on the Nvidia Jetson TX2 embedded platform, our optimized model achieves a 12.7% decrease in inference time compared to the baseline model.

Facial recognition technology plays a crucial role in various domains, including security, surveillance, and biometric authentication. However, the accuracy of conventional facial recognition systems is significantly compromised in degraded conditions, characterized by head pose variations, occlusions, and variations in lighting. In this study, we propose an innovative approach to address this challenge by integrating advanced deep-learning techniques. Specifically, we leverage Cycle Generative Adversarial Networks (CycleGANs) for facial reconstruction and vision Transformers (ViT) for recognition in degraded conditions. GANs are employed to enhance the quality of degraded facial images by generating high-resolution reconstructions from low-quality inputs. The reconstructed images are then fed into Transformer-based models, which extract discriminative features for robust recognition. To evaluate the effectiveness of our approach, we conducted extensive experiments using publicly available datasets of degraded facial images: the EURECOM Kinect Face Dataset and the IST-EURECOM Light Field Face Database. Our results demonstrate significant improvements in face recognition accuracy compared to traditional methods, particularly in scenarios with low-resolution or occluded faces. Furthermore, we analyze the computational efficiency and scalability of our proposed approach, highlighting its potential for real-world deployment.

The human lymphatic system is commonly affected by two primary forms of lymphoma disease: Hodgkin lymphoma and non-Hodgkin lymphoma. The second type, in particular, has emerged as a leading cause of patient mortality. Therefore, achieving a correct and early diagnosis is crucial for healthcare practitioners to devise suitable therapeutic strategies. For these reasons, in this work, We suggest the development of a computer-aided diagnosis system utilizing a novel hybrid approach that incorporates Handcrafted features and BiLSTM networks for the discrimination and analysis of patients with evolving lymphoma from those with residual masses who do not need re-treatment. Our proposed approach combines the concatenation of all extracted features obtained through various handcrafted methods (including histogram analysis, textural, and shape analysis) to analyze the functional, morphological, and anatomical aspects of each lesion. The "LWBDWMRI" databases were utilised for the experiment. We compared the experimental results of the suggested approach to each model: BiLSTM, LSTM, and VGG16. This comparison was conducted across five different approach cases: concatenating functional features only, textural features only, morphological features only, combining textural and morphological features to obtain global anatomical features, and incorporating both functional and anatomical criteria. The proposed approach achieved 96%, 97%, 98%, and 26.11 seconds for Accuracy, F1-score, Recall, and execution time, respectively.

Face recognition systems are essential in various applications. Still, dealing with deteriorated situations such as fluctuations in head posture, lighting, facial expressions, and partial occlusion, presents great difficulty for them. In this work, we provide a novel method for reliable face identification under challenging circumstances by leveraging the Siamese network-based vision transformer architecture. The Siamese network is known for its ability to learn powerful representations from pairs of input data, making it suitable for handling complex variations in face images. We introduce a Transformer-based architecture that integrates Siamese networks to capture long-range dependencies and spatial relationships in facial features effectively. Our method focuses on learning discriminative features from degraded face images, enabling accurate recognition even in challenging conditions. According to experimental findings, our suggested approach works better than current practices in recognizing faces under various degradation factors. The proposed Siamese network-based transformer shows promising results on the two publicly available datasets the EKFD and the IST-EURECOM LFFD offering a reliable solution for face recognition in real-world scenarios with degraded conditions.

Person re-identification is still an open challenging task in various fields due to numerous factors, including illumination changes, background clutter, pose state variations and cloth changes. Several approaches have been suggested to address this problem in the context of deep learning. Generative models, particularly Variational Autoencoders (VAEs), have emerged as promising tools to address these challenges by learning discriminative feature representations of individual images. In this paper, we present Soft-Attention based Person Re-Identification (SAPRI), a novel approach that combines VAEs with a supervised ReID method to enhance the resilience and efficacy of ReID systems. The proposed approach focuses on data reconstruction based on soft attention. Variational autoen-coders encode principally person data, while ignoring irrelevant information. By incorporating supervised ReID, the model learns to appropriately classify persons in real world environments. Our SAPRI proposed method has been evaluated on well-known benchmarks, DukeMTMC-reID and CUHK03, demonstrating superior performance compared to existing state-of-the-art techniques in terms of the mean Average Precision evaluation metric (mAP). Additionally, qualitative results show the effectiveness of the VAE in generating discriminative representations of person images.