Research On Underwater Sonar Image Classification Method Based On Deep Learning

With the rapid development of sonar technology,the classification research of underwater sonar images is deepening.The underwater sonar image classification has important practical significance both in military and civil.In the military,it can help to find the mine,submarine and so on,in the civil,it can help to find the shoal and detect the integrated degree of the dam bottom,which can provide some references for the future underwater object recognition.The traditional classification methods of underwater sonar image adopt different feature extraction methods to complete classification,which causes these methods cannot be widely used.While deep learning models can automatically extract image features through the internal network structure,it has an important influence on image classification.Combining the characteristics of underwater sonar images,learning the theories and methods of deep learning,the deep belief network(DBN)and convolutional neural network(CNN)are constructed for underwater sonar image classification,and the corresponding classification experiments are made.The experimental results show that the classification accuracy of CNN is better than DBN,but the random problem of the initialization of filter weights in CNN will influence the final classification accuracy.To solve the random problem of the initialization of filter weights in CNN,combining the characteristics of underwater sonar images and internal network structure of DBN,a deep learning model with adaptive weights convolutional neural network(AW-CNN)is proposed to classify underwater sonar images.AW-CNN can use the advantage of DBN which can quickly obtain better feature matrix,to adaptively adjust the distribution of filter weights in CNN.The generated weights of DBN are applied to replace the randomly trained filter weights of CNN.The proposed AW-CNN can solve the random problem of the initialization of filter weights in CNN,avoid falling into local optimum,and improve classification accuracy.After the classification of AW-CNN,in order to further improve the classification accuracy of underwater sonar images,the original sonar image dataset is preprocessed(AW-CNN with preprocessed dataset),and finally the classification of underwater sonar image is completed.The experiments with the proposed AW-CNN are from the aspects of effectiveness,convergence and visualization.The proposed AW-CNN has higher classification accuracy under relatively close convergence rate.Meanwhile,the visualization of AW-CNN can intuitively obtain classification results.After verifying the effectiveness,convergence and visualization of AW-CNN,the dataset is preprocessed by a narrowband Chan Vese model with adaptive ladder initialization approach and gray level co-occurrence matrix.The experimental results show that the classification accuracy of AW-CNN with preprocessed dataset is the highest,which proves that dataset preprocessing is conducive to underwater sonar image classification,and also proves the effectiveness of AW-CNN.