Improved Sonar Target Detection Method Based on YOLOv5

Side-scan sonar target detection will soon be used in the field of detect sunken ships and various underwater targets, and has received an increasing attention of scientific researchers in recent years. However, many existing deep-CNNs side-scan sonar detectors still suffer from weaker small target detection performance. It may be attributed to the fact that the layers of a deep convolutional neural network are too deep, which leads to the loss of high frequency information. In this study, we propose a new framework to achieve the robust detection of small targets in side-scan sonar images. Firstly, considering that small targets require more high-frequency features, we propose a Spatial Pyramid Polling Fast++ (SPPF++) module and apply it to the backbone layer network to avoid the loss of feature information for sonar small targets. Secondly, we apply an attention module to the backbone network to improve the recognition ability of small targets. Finally, the Efficient Intersection over Union (EIoU) loss is applied to the regression branch, reducing the network's dependence on the target aspect ratio and improving the accuracy of the regression. Comparing with the state-of-arts detectors, our model has got the better performance. Experiments show that the accuracy of our model has greatly improved compared with YOLOv5-s, and the model parameters have only increased very little, which has practical significance for target detection in side-scan sonar images.