BIDC-YOLO Small target detection in drone aerial photography based on YOLOv8s

Abstract

Small-object detection in unmanned aerial vehicle (UAV) aerial images remains challenging because targets usually occupy only a few pixels and are easily affected by complex backgrounds, occlusion, and illumination changes. To address these problems, this study proposes BIDC-YOLO, an improved YOLOv8s-based detector for UAV small-object detection. The model is redesigned from four aspects. First, a C2f_iSE feature enhancement module is constructed by integrating inverse residual mobile block, squeeze-and-excitation, and efficient multi-scale attention mechanisms into the C2f structure to strengthen spatial and channel feature representation. Second, the original large-object detection branch is replaced with a 160 × 160 small-object detection branch, and bidirectional feature pyramid network is introduced to improve bidirectional cross-scale feature fusion. Third, DySample is used to reduce spatial misalignment during upsampling and preserve fine details of small objects. Finally, cross-layer local attention head is incorporated to enhance local semantic-detail alignment in the detection stage. Experiments on the VisDrone2019 dataset show that BIDC-YOLO improves Precision, Recall, mAP50, and mAP50-95 by 9.6, 10.4, 13.1, and 8.9 percentage points, respectively, compared with YOLOv8s. The results indicate that BIDC-YOLO improves the detection of dense and occluded small objects in UAV aerial scenes.