Research On Object Detection Method Of Side Scan Sonar Based On Deep Learning

With the gradual deepening of people’s exploration of the ocean,the requirements for detection technology are also increasing.However,the marine environment is complex and changeable,and the optical detection methods are greatly limited.The sonar system relies on acoustic imaging and is not affected by factors such as underwater visibility.It has a very important position in the field of underwater detection.At the same time,the image processing technology based on deep learning has developed rapidly in recent years,and has achieved good results in many fields.However,there is still less research on side-scan sonar images using this method.Therefore,this paper studies the target detection algorithm of side-scan sonar images combined with the deep learning method.In view of the lack of side-scan sonar data and the characteristics of sonar images,the labeling method,data expansion method and detection algorithm of side-scan sonar images are studied.The main research contents are as follows:First,the preprocessing and labeling methods of side-scan sonar images are studied.Aiming at the problem that the side-scan sonar image is seriously disturbed by noise,the side-scan sonar image is filtered.Aiming at the problem that the general marking method cannot mark the shadow when marking the side-scan sonar image,a marking method for the side-scan sonar image is designed,which can mark the target and mark the shadow of the target at the same time.The advantages of the improved marking method are demonstrated through a small target generation experiment with a shadowed side-scan sonar.Secondly,a style transfer-based side-scan sonar data expansion method is studied.In view of the problem that the current side-scan sonar image data containing targets is insufficient and cannot meet the training needs of deep learning methods,the style transfer method is used to convert optical images into side-scan sonar images,and for the problem that the original style transfer method cannot well retain the characteristics of the target content and has structural artifacts,the method is improved.The experimental results show that the image generated by the improved method has better effect,which provides data for subsequent target detection research.support.Thirdly,the target detection algorithm based on deep learning is studied,and the YOLOv5 target detection algorithm is selected to detect the side-scan sonar image.In view of the problems that the side-scan sonar image contains less information and the target features are not obvious,the attention mechanism and other modules are added to detect the side-scan sonar image.The algorithm is improved,and the experimental results show that the improved method can effectively improve the detection accuracy of side-scan sonar images.Finally,the small targets in the side scan sonar image are similar to the noise in the image,have no obvious shape features,and are prone to missed detection and false detection.By improving the CenterNet target detection algorithm,the shadow of the target is added for detection.A small target detection method for side-scan sonar is realized,which simultaneously detects the target and the shadow.The shadow information is used to assist the judgment of the target,which improves the detection accuracy of the small target and reduces the missed detection and false detection.