Multi-scale Lightweight Algorithm for UAV Aerial Target Detection

To address the high miss rates and low accuracy in detecting small objects in drone aerial images, caused by limited feature information and noise interference, we propose an improved algorithm, MLA-YOLO, based on YOLOv8s. First, a shallow feature enhancement network with a Global Context Block (GC-Block) is embedded in the backbone to mitigate noise interference and reduce small object feature loss during fusion. Second, deformable convolutions replace part of the standard convolutions in C2F to improve adaptability to geometric variations. Third, an ASPPF module, combined with average pooling, is introduced to enhance multi-scale feature representation and reduce miss rates. Finally, a middle-scale feature synthesis layer is embedded in the neck, with skip connections to ensure smoother transitions between feature scales and enhance feature reuse. Experiments on the VisDrone2019 and VOC2012 datasets show that MLA-YOLO achieves mAP@0.5 values of 40.1% and 72.5%, representing improvements of 8.4% and 3.2% over the baseline YOLOv8s. These results demonstrate the effectiveness and generalization of MLA-YOLO for small object detection in drone images.