Research on The Automatic Detection Method of Side-scan Sonar Image of Small Underwater Target

In order to study the feasibility and correlation performance of applying YOLO model to the automatic detection of side-scan sonar images, based on the complex environment, an underwater scanning test of small targets at sea was designed, and the measured images of side-scan sonar were obtained. After standardized processing, low frequency and high frequency sonar image data sets of underwater small targets were established. Finally, YOLOv5 and YOLOv7 were used as automatic detection models to train and test the acoustic atlas. The results show that the accuracy and recall rates of YOLOv5 and YOLOv7 models on the two acoutomograms can both reach more than 60%, which verifies the feasibility of using YOLOv5 and YOLOv7 models for automatic detection of underwater small targets by side-scan sonar. Further analysis of the experimental results shows that the accuracy rate and recall rate are different in different combinations of models and data sets, indicating that different combinations of detection models and atlas can only perform well in a certain index, and scanning activities should flexibly select detection models according to task requirements and target characteristics. For metal targets, various combinations of detection models and atlas show that the detection performance of low-frequency sonar images is better than that of high-frequency images, which changes the intuitive understanding that high-frequency images are clearer than low-frequency images in traditional human eye image judgment, and is of great significance for improving the detection range and detection and recognition rate of side-scan sonar.