Salma Kammoun Jarraya

Deep learning has enabled significant progress in source code generation, aiming to reduce the manual, error-prone, and time-consuming aspects of software development. While many existing models rely on recurrent neural networks (RNNs) with sequence-to-sequence architectures, these approaches struggle with the long and complex token sequences typical in source code. To address this, we propose a grammar-based convolutional neural network (CNN) combined with a tree-based representation to enhance accuracy and efficiency. Our model achieves state-of-the-art results on the benchmark HEARTHSTONE dataset, with a BLEU score of 81.4 and an Acc+ of 62.1%. We further evaluate the model on our proposed dataset, AST2CVCode, designed for computer vision applications, achieving 86.2 BLEU and 51.9% EM. Additionally, we introduce BLEU+, an enhanced evaluation metric tailored for functional correctness in code generation, which achieves a BLEU+ score of 92.0% on the AST2CVCode dataset. These results demonstrate the effectiveness of our approach in both model architecture and evaluation methodology.

Deepfake video detection has emerged as a critical challenge in the realm of artificial intelligence, given its implications for misinformation and digital security. This study evaluates the generalisation capabilities of the CoAtNet model—a hybrid convolution–transformer architecture—for deepfake detection across diverse datasets. Although CoAtNet has shown exceptional performance in several computer vision tasks, its potential for generalisation in cross-dataset scenarios remains underexplored. Thus, in this study, we explore CoAtNet’s generalisation ability by conducting an extensive series of experiments with a focus on discovering features and variations in deepfake videos. These experiments involve training the model using various input and processing configurations, followed by evaluating its performance on widely recognised public datasets. To the best of our knowledge, our proposed approach outperforms state-of-the-art models in terms of intra-dataset performance, with an AUC between 81.4% and 99.9%. Our model also achieves outstanding results in cross-dataset evaluations, with an AUC equal to 78%. This study demonstrates that CoAtNet achieves the best AUC for both intra-dataset and cross-dataset deepfake video detection, particularly on Celeb-DF, while also showing strong performance on DFDC.

Road object segmentation is crucial for autonomous driving, as it enables vehicles to perceive their surroundings. While deep learning models show promise, their generalization across diverse road conditions, weather variations, and lighting changes remains challenging. Different approaches have been proposed to address this limitation. However, these models often struggle with the varying appearance of road objects under diverse environmental conditions. Foundation models such as the Segment Anything Model (SAM) offer a potential avenue for improved generalization in complex visual tasks. Thus, this study presents a pioneering comprehensive evaluation of the SAM for zero-shot road object segmentation, without explicit prompts. This study aimed to determine the inherent capabilities and limitations of the SAM in accurately segmenting a variety of road objects under the diverse and challenging environmental conditions encountered in real-world autonomous driving scenarios. We assessed the SAM’s performance on the KITTI, BDD100K, and Mapillary Vistas datasets, encompassing a wide range of environmental conditions. Using a variety of established evaluation metrics, our analysis revealed the SAM’s capabilities and limitations in accurately segmenting various road objects, particularly highlighting challenges posed by dynamic environments, illumination changes, and occlusions. These findings provide valuable insights for researchers and developers seeking to enhance the robustness of foundation models such as the SAM in complex road environments, guiding future efforts to improve perception systems for autonomous driving.

Transforming images from day style to night style is crucial for enhancing perception in autonomous driving and smart surveillance. However, existing CycleGAN-based approaches struggle with texture loss, structural inconsistencies, and high computational costs. In our attempt to overcome these challenges, we produced LBP-CycleGAN, a new modification of CycleGAN that benefits from the advantages of a Local Binary Pattern (LBP) that extracts details of texture, unlike traditional CycleGAN, which relies heavily on color transformations. Our model leverages LBP-based single-channel inputs, ensuring sharper, more consistent night-time textures. We evaluated three model variations: (1) LBP-CycleGAN with a self-attention mechanism in both the generator and discriminator, (2) LBP-CycleGAN with a self-attention mechanism in the discriminator only, and (3) LBP-CycleGAN without a self-attention mechanism. Our results demonstrate that the LBP-CycleGAN model without self-attention outperformed the other models, achieving a superior texture quality while significantly reducing the training time and computational overhead. This work opens up new possibilities for efficient, high-fidelity night-time image translation in real-world applications, including autonomous driving and low-light vision systems.

Improving the recognition of online learning engagement is a critical issue in educational information technology, due to the complexities of student behavior and varying assessment standards. Additionally, the scarcity of publicly available datasets for engagement recognition exacerbates this challenge. The majority of existing methods for detecting student engagement necessitate significant amounts of annotated data to capture variations in behaviors and interaction patterns. To address these limitations, we investigate few-shot learning (FSL) techniques to reduce the dependency on extensive training data. Transformer-based models have shown comprehensive results for video-based facial recognition tasks, thus paving new ground for understanding complicated patterns. In this research, we propose an innovative FSL model that employs a prototypical network with the vision transformer (ViT) model pre-trained on a face recognition dataset (e.g., MS1MV2) for spatial feature extraction, followed by an LSTM layer for temporal feature extraction. This approach effectively addresses the challenges of limited labeled data in engagement recognition. Our proposed approach achieves state-of-the-art performance on the EngageNet dataset, demonstrating its efficacy and potential in advancing engagement recognition research.

In recent decades, unmanned aerial vehicles (UAVs) have become widely utilized for many real-world applications, including surveillance, crowd management, and threat detection, providing a new perspective to recognize human behaviors. However, current UAV-based video datasets adopt categorization schemes that rely on broad and inconsistent categories relative to real-world aerial contexts. To address this knowledge gap, this study proposes a novel human activity categorization framework derived from a comprehensive systematic analysis study of ten publicly available UAV-based human action recognition (HAR) datasets, incorporating a variety of environmental situations and human behaviors. By reconciling inconsistent categories and finer activities, this taxonomy serves as a standard framework for UAV-based HAR research. The proposed categorization framework is validated by comparing it with other existing frameworks on the publicly benchmarked Drone-Action dataset, outperforming them by 97% across four metrics. Our contribution aims to develop the foundation for further experimental validation and provide a guide for researchers interested in developing accurate and context-aware surveillance systems.

Drug abuse and addiction problems are one of the most serious health, social, and psychological problems facing the world. Many international studies indicate that the start of drug abuse occurs mostly in adolescence, which is the period that young people spend in schools, institutes, and universities. Drugs in the student community have become a scourge that raises increasing concern, whether among families or educators, over the fate of school children and educational attainment. Regarding their behaviors, an addicted student often exhibits abnormal behaviors such as permanent lethargy, anxiety, tremors, and aggressive behavior toward others. Moreover, to obtain drugs, the addicted student becomes compelled to resort to various means and ways, and they gradually become criminal addicts. To this endeavor, a detector of abnormal behaviors in schools has become a necessity. In this paper, we built an automatic system able to analyze and detect abnormal behaviors of addicted students and inform the educational staff and parents to know how to manage and treat them. On a technical level, we used deep learning and the recent computer vision techniques in the suggested solution due to their contributions to human behavior and emotion recognition fields. The best‐recorded result (97.5%) is obtained with fused handcrafted features based on skeleton joints and deep features extracted with the MobileNet pretrained model and forwarded to a deep proposed network based on two TimeDistributed layers, one BiLSTM layer, and several Dense layers.

This paper explores the efficacy of Active Learning (AL) techniques, specifically focusing on Dynamic Data Selection (DDS), for improving image classification tasks. AL is a machine learning paradigm that enables the automatic selection of the most informative data samples for annotation, thereby reducing the annotation burden and enhancing model performance. In this study, we investigate the integration of DDS techniques with AL strategies to iteratively select the most informative image samples for model training. We use a fine-tuned VGG16 as the underlying classification model due to their effectiveness in image analysis tasks. Our experimental evaluation involves comparing the performance of fine-tuned VGG16 trained with three AL-based DDS techniques on Arabic sign language dataset. We analyze various DDS strategies, including Random selection, Entropy-based selection, and margin selection to determine their impact on model accuracy and annotation efficiency. The results of our study demonstrate the effectiveness of margin selection method-based AL approach in improving the performance of recognition of 32 hand gestures for Arabic sign language (95.3 \%) while minimizing the annotation effort.

Software reliability is a crucial factor in determining software quality quantitatively. It is also used to estimate the software testing duration. In software reliability testing, traditional parametric software reliability growth models (SRGMs) are effectively used. Nevertheless, a single parametric model cannot provide accurate predictions in all cases. Moreover, non-parametric models have proven to be efficient for predicting software reliability as alternatives to parametric models. In this paper, we adopted a deep learning method for software reliability testing in computer vision systems. Also, we focused on critical computer vision applications that need high reliability. We propose a new deep learning-based model that is combined and based on the ensemble method to improve the performance of software reliability testing. The experimental results of the new model architecture present fairly accurate predictive capability compared to other existing single Neural Network (NN) based models.