Research On Underwater Target Recognition Technology Based On Deep Neural Network

With the depletion of terrestrial resources caused by daily consumption of human beings,the ocean will become one of the important sources of human resources in the future.Underwater target recognition technology is an indispensable part of human exploration of marine resources,and it has attracted more and more attention from researchers in various countries.In recent years,with the success of deep learning technology in various fields,this paper improves the deep neural network model by studying the commonly used deep neural networks(such as Convolutional Neural Networks and Recurrent Neural Networks)algorithms,combined with traditional machine learning knowledge and other parameter optimization methods.Using deep neural network to extract features to replace existing traditional extraction features(such as wavelet transform,Mel-frequency Cepstrum Coefficient and Hilbert-Huang transform),explore the feasibility of target recognition under underwater scenes In view of the higher recognition accuracy than the underwater target recognition method based on traditional feature extraction.Specifically,this article works as follows:1)Several typical feature extraction and classification methods are introduced in detail as the benchmark method for the comparative experiments in this paper.In addition,the basic theory of the two network models of the current deep neural network is expounded.2)By analyzing the advantages and disadvantages of the commonly used loss functions,an underwater target recognition method based on Convolutional Neural Networks(CNN)is proposed.This method embeds a regular term into the basic mean square error loss function to obtain the objective function.At the same time,based on the improved objective function,this paper analyzes the forward propagation and back propagation of CNN,and improves the CNN of underwater target recognition which is suitable for the data set of this paper.3)By analyzing the advantages and disadvantages of the Recurrent Neural Network(RNN)and the Long Short-Term Memory networks(LSTM),an underwater target recognition method based on LSTM-RNN is proposed.This method combines the characteristics of the dataset of this paper.By using LSTM hidden layer neurons to replace the native RNN hidden layer neurons,LSTM-RNN is established,which avoids the influence of gradient explosion and gradient disappearance caused by long-term dependence of RNN.The two methods proposed in this paper have been verified by the data set of active sonar echo signals,and all have achieved good results.Compared with the traditional machine learning method and the improved CNN method,the LSTM-RNN method improves the target recognition accuracy of underwater scenes by at least 2%.