Small Objects Recognition by Exploiting an Improved YOLOv5 Algorithm on the UAV Platform

The aerial imagery captured by unmanned aerial vehicles (UAVs) exhibits several challenging characteristics such as large variations in object size and complex backgrounds, which make it difficult for existing detectors to detect small objects in such imagery. To address the problem of false or missed detections of small objects in UAV imagery, the BHF-YOLOv5 detection model is proposed. Firstly, the Biformer module is added to the backbone network, which employs sparse sampling to retain fine-grained details in the image while reducing computation, thereby enhancing attention to small objects and effectively improving the problem of small object detection. Secondly, a high-resolution detection head is added to the detection layer, which effectively increases the network's sensitivity to small objects. By combining multi-scale detection and multi-level information fusion, the network's representation ability is significantly enhanced. Furthermore, the Focal EIoU Loss optimization loss function is used to address the problem of imbalanced samples, thereby optimizing the network's training efficiency. Multiple experimental results on the VisDrone dataset show that our proposed BHF-YOLOv5 model improves the mAP metric by 3.19% compared to the original YOLOv5 model. Compared with other mainstream models, our proposed model achieves better detection performance for small objects in UAV imagery and can effectively complete the small object detection task of UAVs.