BRA-YOLO: Object Detection Algorithm with Bi-Level Routing Attention for YOLOv5

At present, YOLOv5 is the most popular algorithm in single-stage target detection, has covered all areas of society. However, because the neck layer can not effectively integrate the context information content, it is still difficult to identify the small target features incorrectly and omitted. In addition, YOLOv5 also faces the problem of low detection accuracy. In response to the above issues, in this paper, BiFormer attention mechanism is introduced into the Neck C3 module of YOLOv5s model. Based on the above issue, this paper introduces the C3 module with the BiFormer attention mechanism into the Neck of the YOLOv5s model, proposing the BRA-YOLO object detection algorithm. This algorithm aims to optimize the detection ability of the model for small target features and improve the detection accuracy of the model. Experimental results show that compared with YOLOv5s, the BRA-YOLO target detection algorithm has a significant improvement of 1.303% on mAP@0.5%, 1.389% on mAP@0.5:0.95%, and 2.32% on small target detection in general. BRA-YOLO effectively improves the overall accuracy of the model and the accuracy of small target detection.