Target Detection In Colorful Imaging Sonar Based On Convolutional Neural Network

When divers explore and salvage in an underwater environment,they can only find the approximate location of the target through sonar imaging.It makes divers who are often in complex underwater environment have greater risks when they work underwater.Therefore,it is expected to realize automatic underwater target detection with the use of underwater robots in the future.Target detection has evolved from the traditional methods of feature extraction and classifier classification detection to today’s application of convolutional neural network.The research of convolutional neural networks and the emergence of various algorithms have greatly improved the efficiency and accuracy of target detection,which lays a foundation for the color imaging sonar target detection in this paper.Compared to the popular public dataset,the sonar imaging data set is relatively rare.The TKIS-I helmet-type color image sonar developed by our laboratory has been approved by the Chinese Naval Navigation Support Department.At present,there are more than 20 units serving the troops and continuously obtaining orders from the troops.The imaging results of the sonar can be used as a data collection tool.Combining the characteristics of sonar imaging and the target detection algorithm based on neural network,this paper uses the real-time YOLOV3 as the basic network under the condition of ensuring high detection accuracy,and employs the bilinear interpolation algorithm for the pretreatment of the sonar target image.In addition,the YOLOv3 network is improved by the target prediction frame dimension clustering,the selection of the optimal prediction frame number,the optimal threshold selection,the different resolution data optimization network,the multi-scale training,and the improved detection network to make it better for color image sonar target detection.The author has completed the following work:1.Outline the current situation and development of underwater detection in China and foreign countries,and the research status of imaging sonar target detection.2.Summarize the development,application and structure of convolutional neural networks.Summarize the structure and characteristics of the YOLOv3 network feature extraction layer and detection layer.Analyze the reasons for using the activation function and the loss function in the network,and compare other existing network models to explain the reason why YOLOv3 is selected as the basic network model.3.Realize the whole process of image target detection of YOLOv3 network applied to imaging sonar.This includes the configuration of the experimental environment and the construction of the experimental platform;the main components and imaging characteristics of the TKIS-I helmet-type color image sonar developed by the laboratory;the pre-processing of the imaged sonar image according to the characteristics of sonar imaging;Sonar target data set.lastly,the author uses the original YOLOv3 network for training and testing,but the effect is not ideal,which paves the way for the improvement of the next chapter.4.Improve color imaging sonar target detection algorithm.The author improves the YOLOv3 network mainly from five aspects:the target prediction frame dimension clustering,optimal threshold selection,different resolution data optimization network,multi-scale training,and improved detection layer network.So it can train the ideal model for the detection of sonar targets.5.Evaluate the detection performance and effectiveness of the target detection algorithm before and after improvement.Firstly,the evaluation indexes commonly used in target detection are introduced.Then,the experimental results of each improved method in the previous chapter are compared and analyzed.Finally,the improved new network is compared with the experimental results of other target detection algorithms to obtain the optimal effect picture of target detection.6.Summarize the results and shortcomings of the image sonar target detection technology designed in this paper,and put forward some aspects that need to be explored and improved in the future.