Sonar Image Target Recognition And Detection Based On Convolutional Neural Network

With the maturity of acoustic imaging technology and the development of high resolution imaging sonar,underwater target recognition and detection using sonar images has become an important means of coastal defense and Marine resources exploration.In view of the shortcomings of traditional sonar image target recognition and detection methods,such as complicated steps,low efficiency and relying on human experience,this paper studies the sonar image target recognition and detection methods based on convolutional neural network.The main contents and achievements are as follows:(1)The basic concept and principle of convolutional neural network(CNN)are introduced in detail,and the target recognition and target detection methods based on CNN are expounded.(2)In order to study what kind of CNN network structure has better effect on sonar image target recognition,we design five different CNN models,and study the effect of different network layers,convolutional kernel size and activation function on sonar image recognition results.(3)In order to compare CNN method with traditional methods of target recognition,in this paper,we simply introduce the principle of traditional target recognition method based on artificial characteristic and analyze two kinds of common feature extraction of image target recognition field operator: the basic principle of HOG operator and LBP operator.We introduce the support vector machine(SVM)classifier classification process and study the effect of different SVM kernel functions on sonar image target recognition results.(4)The CNN based YOLOv3 algorithm is applied to the sonar image target detection task,and the k-means clustering algorithm was used to improve the performance of YOLOv3 algorithm.(5)In view of the shortcomings of YOLOv3 in sonar image target detection,such as computational redundancy and insufficient feature fusion,an improved FW-YOLOv3 target detection algorithm model is proposed in this paper.Compared with YOLOv3,our FWYOLOv3 algorithm achieve higher accuracy and faster detection speed in sonar image target detection task.The innovation points of this paper are summarized as follows:1.In sonar image target recognition task,5 different CNN network models are designed,and the effects of different network layers,convolution kernel size and activation function on sonar image recognition results are compared in the experiment.2.In sonar image target detection task,an improved FW-YOLOv3 algorithm model is proposed.In this new algorithm,we design a single-scale target prediction network and propose a new weighted feature fusion algorithm.The experimental results show that the proposed single-scale prediction network can improve the detection speed and the weighted feature fusion algorithm can effectively improve the detection accuracy of the model.