Mohand Saïd Allili

Emerging deep learning architectures and large foundation models have transformed medical image segmentation, yet significant challenges remain when limited annotations are available across diverse imaging modalities and anatomical structures. Current few-shot learning approaches struggle with limited training examples and fail to leverage anatomical expertise effectively, often relying on superficial visual similarities rather than meaningful anatomical correspondences essential for medical image analysis. We propose FUSE-RAG, a novel framework based on retrieval-augmented generation (RAG) for few-shot medical image segmentation. For an image query, FUSE-RAG uses a retrieved anatomically-relevant support set for conditioning mask generation. Our key innovations include: (1) an ROI-aware retrieval mechanism that injects expert anatomical knowledge into medical foundation model features, guiding them toward clinically relevant regions and aligning with core RAG principles for medical imaging, (2) a segmentation architecture integrating Anatomical Correspondence Blocks (ACB) with Vision Mamba SS2D for efficient long-range modeling, Support Quality Assessment Blocks (SQAB) for adaptive feature weighting, and Support-Conditioned Skip Connections (SCSC) for propagating anatomical guidance. FUSE-RAG extends prompt engineering principles from natural language processing to medical imaging, demonstrating that carefully selected anatomically relevant examples outperform larger randomly chosen sets through a quality-over-quantity paradigm. Comprehensive evaluation on four standard datasets, ATLAS 2.0 stroke lesion, QaTa-COVID19 pneumonia, ISIC 2018 skin lesion, and DRIVE retinal vessel, demonstrates substantial improvements of 12.69%, 10.31%, 5.60%, and 11.79% Dice coefficient, respectively, over state-of-the-art few-shot methods, with the potential of real-time deployment. Code will be made publicly available at: https://github.com/MOHAMEDLamine77/FuseRAG.

Identifying safe landing zones (SLZ) for Unmanned Aerial Vehicles (UAVs) is important to ensure safe navigation, especially when operated in critical environments. However, this is a challenging task, especially in dynamic scenes characterized by concurrent movement of obstacles and UAVs. This paper proposes a vision-based framework that maps SLZs in dynamic scenes, characterized by both object and drone motions, by integrating several functionalities for visual analysis of the UAV surrounding scenes. Static scene analysis is achieved through context-aware segmentation which divides the UAV-acquired images into thematic classes enabling to identify suitable landing surfaces (e.g., roads, grass).For dynamic content analysis, we combine object detection, tracking, and trajectory prediction to determine object occupancy and identify regions free of obstacles. Trajectory prediction is performed through a novel encoder-decoder architecture taking past object positions to predict the most likely future locations. To ensure stable and robust trajectory prediction, we introduce an optimized homography computation using multi-scale image analysis and cumulative updates to compensate UAV motion. We tested our framework on different operational scenarios, including urban and natural scenes with moving objects like vehicles and pedestrians. Obtained results demonstrate its strong performance, and its significant potential for enabling autonomous and safe UAV navigation.

Feature extraction involves the identification and extraction of salient characteristics or patterns, including edges, textures, shapes, and color attributes. Contemporary feature extractors predominantly leverage deep learning architectures, such as Convolutional Neural Networks (CNNs) and Vision Transformers (VITs). The availability of diverse feature extractors in the literature provides a wide range of feature representations. Features extracted from an image depend on the specific application, the chosen extractor, and its configuration. Therefore, integrating complementary information by combining distinct extractors offers a promising way to enhance performance. Graph Neural Networks (GNNs), particularly Graph Convolutional Networks (GCNs), have emerged as powerful and widely adopted approaches for semi-supervised image classification, as they effectively leverage both labeled and unlabeled data while exploiting the underlying graph structures that capture relationships among samples. This study proposes a novel approach for GNNs in scenarios where labeled data is scarce, by integrating diverse sets of feature and graph representations derived from various extractors in classification scenarios. Experimental investigations were conducted, encompassing combinations of distinct feature and graph extractors, as well as rank aggregation strategies. The primary contributions of this work are underscored by the experimental findings, which demonstrate that the strategic combination of feature and graph representations, coupled with the application of manifold learning for graph processing, leads to significant improvements in classification accuracy across the majority of experimental conditions. Furthermore, the utilization of rank aggregation techniques to integrate features from different extractors was shown to enhance classification accuracy.

As Unmanned Aerial Vehicles (UAVs) see growing use in civilian applications, reliably identifying Safe Landing Zones (SLZs) in varied environments is essential for autonomous navigation and emergency response. Passive vision sensors offer a low-cost, lightweight solution for real-time terrain analysis and 3D scene reconstruction, making them ideal for onboard systems. We introduce OR-SLZNet, an original deep learning model based on ordinal regression to predict SLZs from UAV imagery. Unlike prior approaches, OR-SLZNet produces dense, multi-level safety maps by jointly leveraging photometric (e.g., color and texture) and geometric cues (e.g., flatness, slope, and depth), assigning each pixel an ordinal safety score that reflects landing suitability. With real-time inference (~0.02s/frame), the model supports onboard deployment and rapid decision-making in time-critical situations. Extensive experiments on five diverse datasets demonstrate OR-SLZNet effectiveness and strong generalization across a wide range of structural complexities. Additional materials and code are available at https://github.com/saba129/OR-SLZNet.

DeepFake technology has advanced significantly in recent years, enabling the creation of highly realistic synthetic face images. Existing DeepFake detection methods often struggle with pose variations, occlusions, and artifacts that are difficult to detect in real-world conditions. To address these challenges, we propose a multi-view architecture that enhances DeepFake detection by analyzing facial features at multiple levels. Our approach integrates three specialized encoders, a global view encoder for detecting boundary inconsistencies, a middle view encoder for analyzing texture and color alignment, and a local view encoder for capturing distortions in expressive facial regions such as the eyes, nose, and mouth, where DeepFake artifacts frequently occur. Additionally, we incorporate a face orientation encoder, trained to classify face poses, ensuring robust detection across various viewing angles. By fusing features from these encoders, our model achieves superior performance in detecting manipulated images, even under challenging pose and lighting conditions.Experimental results on challenging datasets demonstrate the effectiveness of our method, outperforming conventional single-view approaches