An Efficient and Lightweight Framework for Small Object Detection in UAV Images

Detecting small objects in unmanned aerial vehicle (UAV) images is challenging due to complex backgrounds and insufficient feature representation. This paper proposes feature enhance and scale aware detector (FESA-DET), an efficient and lightweight framework for UAV object detection. It introduces a dual feature enhance module (DFEM) that uses parallel local and global perception unit with adaptive fusion to capture richer multi-scale context, enhancing feature richness and adaptability for small objects. Additionally, a feature scale aware module (FSAM) employs spatial and channel aware mechanism recalibrate features, mitigating scale confusion by suppressing large object features and enhancing small object details. Experiments on VisDrone and UAVDT benchmarks show FESA-DET achieving mAP50 of 49.3%, 32.0% and mAP50-95 of 30.2%, 17.6% respectively, demonstrating advanced performance over current methods. With only 1.27M parameters and 41.7 GFLOPs, FESA-DET is an efficient, high-performance detector for small objects in UAV images.