Research On Deep Learning-based Underwater Target Recognition Method

At present,there are two research directions in underwater target recognition application:one direction is based on passive detection,in which the researhcers extract underwater sound?noise?feature before underwater target noise identification;the other direction is using imaging sonar to get underwater acoustic images,and then the classifier is used to recognize the underwater target throw the image processing and feature extraction.However,the method of passive detection often results in poor recognition because of weak signal,but the active detection method can often obtain higher echo intensity.Manually extracting the features of underwater acoustic images or extracting underwater target acoustic features will inevitably lose some of the key information.Using the method of deep learning,underwater sonar recognition of the original sonar echo information can avoid the problem of losing the information by artificial features.The main recognition target of this thesis is the active sonar detection echoes.Time series form samples and time-frequency images are used as data sets to explore the application of deep learning method in underwater target recognition.Firstly,the modeling method and optimization of target function in deep learning are analyzed,including the calculation and comparison of loss functions,the comparison and selection of nonlinear activation functions,the analysis and improvement of parameter optimization methods and the enumeration and application of partial regularization methods.Secondly,the calculation principle,network structure and modeling method of multilayer perceptrons,convolutional neural networks and recurrent neural networks in the field of target recognition are discussed.Based on the previous studies,this thesis considers the multilayer perceptron with bottleneck as a single multi-layered perceptron structure,together with the structures of"straight""linear""exponential"as four network structures types of multilayer perceptron.For dimension disaster of deep learning,the unsupervised neural network dimension reduction method of automatic encoder and restricted Boltzmann machine are combined with the multilayer perceptron in the task of deep learning target recognition to provide a new idea of underwater target recognition based on dimension reduction.Finally,this thesis evaluates the performance of each model in terms of convergence error,F₁ value,ROC,AUC and confusion matrix of neural network.The recognition result of deep learning model with 100-50-100 hidden layer structure is selected from 16 kinds of multilayer perceptrons to compare with multilayer perceptron combined with the automatic encoder,in which the original eigenvector is reduced to a three-dimensional,and the multilayer perceptron combined with restricted Boltzmann machine,in which the original eigenvector is reduced to a five-dimensional,and the convolution neural network without pooling layer with 12-24-24 distribution of convolution kernel.According to the discussion and research analysis above,convolution neural network model with convolution kernel distribution of 12-24-24 is more suitable for underwater target recognition in this thesis.The recall rate of underwater frogman,empty target,and oxygen bottle reached 100%,92%and 63%,AUC are 1.000,0.9374,and0.9409,which get better recognition result for underwater threat targets.