DroneScan-YOLO: Redundancy-Aware Lightweight Detection for Tiny Objects in UAV Imagery

Researcher Yann V. Bellec has introduced DroneScan-YOLO, a novel computer vision architecture specifically engineered to solve the persistent challenge of detecting tiny objects in drone-captured imagery. The system addresses three critical limitations of standard YOLO detectors: their inability to detect objects smaller than 32 pixels, gradient vanishing issues with non-overlapping bounding boxes, and architectural redundancy. Through four coordinated innovations, including a higher 1280x1280 input resolution and a new P2 detection branch at stride 4 (adding just 114,592 parameters), the model fundamentally improves spatial detail capture for minute targets.

On the VisDrone2019-DET benchmark, DroneScan-YOLO delivers a substantial leap in performance, scoring 55.3% mAP@50 and 35.6% mAP@50-95. This represents gains of +16.6 and +12.3 points respectively over the YOLOv8s baseline, while crucially maintaining a real-time inference speed of 96.7 FPS. The improvements are most dramatic for the smallest object categories, with bicycle detection AP@50 soaring by 187% from 0.114 to 0.328.

The architecture's efficiency stems from its RPA-Block, a dynamic filter pruning mechanism that reduces computational waste, and its novel SAL-NWD loss function. This hybrid loss combines Normalized Wasserstein Distance with size-adaptive CIoU weighting, integrated into YOLOv8's TaskAligned assignment pipeline, to provide stable gradients for tiny, non-overlapping boxes. The result is a practical, lightweight model that doesn't sacrifice speed for accuracy, making advanced aerial surveillance and analysis feasible on edge devices.