BCF-YOLO: A Real-Time Detection Algorithm for Small Targets in Remote Sensing Images Based on Regional Information Sensing

In order to solve the dilemma of detecting small targets in remote sensing images, this paper proposes a remote sensing image small target detection algorithm BCF-YOLO based on regional information fusion and multiple attention mechanisms based on YOLO v5. First, we propose a BAT module (interactive regional attention module) in the backbone, which fuses spatial and channel information in a parallel interactive way, and adopts a two-layer routing dynamic sparse attention to realize a more flexible and content-aware target detection algorithm. sparse attention to realize more flexible computational allocation with content-awareness. Second, the CF module (region-aware, information fusion module) is proposed in the feature fusion stage to perform feature fusion with reuse of feature information and close attention to the contextual information of key targets. Finally, an additional small target detection head is added and Wise IoU is introduced, while the C3 module is improved to ensure real-time monitoring with lower computational complexity and to improve the practicality of the whole detection model. The experimental results show that the recognition rate of mAP0.5:0.95 on the VisDrone2019 dataset is improved from 31.5% to 34.9% compared with YOLO v5s. Compared with other target detection algorithms, BCF-YOLO shows superior performance in small target detection on remote sensing images.