Target detection in color sonar image based on YOLOV5 network

Sonar is the main equipment used to explore underwater. The traditional sonar target detection technology has the problems of low intelligence, poor robustness, poor real-time performance and low recognition accuracy. In this paper, a high-precision target detection algorithm named RFB-SE-YOLOV5 is proposed. The data set was produced by sonar made in the laboratory based on the YOLOV5 network model. The SE attention mechanism was added to the original backbone network to improve the convergence of the model and the ability to extract effective features of models. Besides, the RFB-SE module is addedto the backbone network to increase the Receptive Field of the network, thereby enhancing the model for differentiability and robustness characteristics. RFB-SE module is the attention module added after concat of RFB (Receptive Field Block). After testing 660 sonar images, the mAP@0.5 of the improved RFB-SE- Yolov5 network is 0.972, which is 3.2% higher than the original network. The improved algorithm can be applied to underwater robot automatic detection of underwater targets in the future.