DRPE-YOLO: a multiscale feature fusion network for small target detection in aerial images

Autonomous landing of drones heavily relies on accurate recognition of small ground targets from visual information. However, existing methods face challenges such as low contrast, occlusion, and dense distribution of small targets in aerial images. To address these challenges, we proposed DRPE-YOLO, a small target detection algorithm specifically optimized for drone autonomous landing. First, a residual feature enhancement module is integrated into the backbone network. It enhances the capture of contextual information via dilated convolutions. Second, in the neck section, a deep feature pyramid network incorporating depthwise separable convolution is designed. This effectively reduces spatial and channel redundancy while expanding the sampling depth. Third, a partial convolution enhanced module, which combines partial convolution and cross-spatial attention, is proposed to improve feature aggregation for small targets. Extensive experiments conducted on the AI-TOD, DIOR, and VisDrone2019 datasets demonstrate the superiority of DRPE-YOLO. Compared with YOLOv5, our method achieves significant improvements on the AI-TOD dataset: AP50 is increased by 15% (54.4% versus 39.4%), and AP50-95 is improved by 5.6% (21.2% versus 15.6%). Furthermore, DRPE-YOLO demonstrates strong generalization ability, achieving AP50 gains of 4.7% on the DIOR dataset and 11% on the VisDrone2019 dataset. These results indicate that the proposed framework significantly promotes the development of small target detection in aerial scenarios.