Research On Target Detection Method Of Side-scan Sonar Images Based On Deep Learning

The side-scan sonar is the most widely used and most important equipment in the underwater target recognition task and submarine geomorphologic information survey.However,due to the low resolution of side-scan sonar imaging,serious noise and image background interference,some targets that are very similar to the background and overlapped small targets are extremely difficult to distinguish.These interference factors make the target detection task pretty difficult.In this paper,aiming at the above difficulties,four kinds of targets with different mesoscale of side-scan sonar images are detected based on the deep learning method.The specific research contents are as follows.Firstly,the imaging characteristics of side-scan sonar are briefly introduced,then the research status of the target detection technology at home and abroad and the existing problems of underwater target recognition technology are described in detail.The difficulties and challenges in the target detection of the existing domestic outboard scanning sonar image are also analyzed.Facing the problem of low accuracy and poor generalization ability of the current underwater side-scan sonar target detection method,this paper chooses the method based on deep learning to realize the accurate detection of the side-scan sonar image target.Secondly,the enhancement and preprocessing of side-scan sonar image data are completed,and the data set of side-scan sonar images is established.Due to the high acquisition cost of side-scan sonar images,the amount of available data is very scarce,and the image after imaging is seriously affected by noise and background interference.In order to solve these problems,some commonly used data enhancement methods such as geometric transformation is proposed,and some side-scan sonar images containing targets are generated by GAN to complete the expansion of the data set,which not only obtains high-quality data,but also lays a good data foundation for the application of deep learning algorithm.Thirdly,the target detection algorithm of side-scan sonar image based on improved YOLOv3 and transfer learning is studied.The basic structure of the original YOLOv3 algorithm is analyzed,and then an improved YOLOv3 algorithm model is proposed to enhance the feature connection at network level and enrich the feature layer by combination pooling,aiming at the missing and misdetection problems and unique shadow features in the detection of side-scan sonar images.Furthermore,the generalization ability of the improved model is increased by the transfer learning method,which effectively improves the detection accuracy of the target.Finally,the real-time detection model based on the improved center point algorithm is completed,and the detection effect of overlapping small targets is enhanced by combining with the method of multi-scale testing.Aiming at the problem that there is more information in the side-scan sonar image and the slow detection speed,an improved detection algorithm of the center point is designed,which is not dependent on the anchor boxes.In addition,on the basis of the improved model,the method of multi-scale testing is combined to solve the problem of insufficient detection accuracy of overlapping small targets.In this paper,the improvement of accuracy and speed of the improved algorithm is proved through multiple comparison experiments,and the problems of missing detection,misdetection and low detection accuracy of overlapping small targets in the detection process of side-scan sonar image are solved.Finally,this paper points out some shortcomings and predicts the future application of deep learning in the detection of side-scan sonar images.