Research On The Optimization Method Of Convolutional Neural Network For Underwater Target Recognition

With the development of the times and the science and technology,land resources have gradually failed to meet human needs,and human beings have begun to turn to the exploitation and development of ocean resources.All the countries of the world have also erupted many military conflicts in maritime because of the related competition of ocean resources.The economic and military importance of the ocean is becoming more and more prominent.Therefore,the study of underwater target recognition has profound significance in two aspects of economy and military.Due to the increasing complexity of underwater sound signals,the limitations of some traditional target recognition methods,including classical statistics and artificial recognition,are becoming more and more obvious.Therefore,how to realize underwater target recognition accurately and efficiently has become an important direction of current research.At present,deep learning based on bionic neural network has achieved many good results in target recognition field,especially convolutional neural network has surpassed human recognition ability in the field of image and sound.Therefore,the purpose of this thesis is optimizing the convolutional neural network in the underwater target recognition field.In the study of underwater target recognition's work,the convolutional neural network has a better recognition ability of image signal,this thesis attempts to change an underwater sound signal into an image signal,and trains and predicts by using a convolutional neural network model.The composition of traditional convolutional neural network is mainly composed of two parts: feature extraction and classification.Therefore,some optimizations are proposed by this thesis from above two perspectives to achieve better results in underwater target recognition.In this thesis,considering the pixel correlation of spectrum image,full convolution is proposed to extract the 1×1 feature map,which combined with the level of the acoustic image.A feature extraction method is proposed based on multiscale filter banks;in order to avoid the decrease of the precision in the convolutional neural network,a residual learning method is introduced into the model in this thesis.Considering the stability and progressiveness of the Alexnet network,an optimized model based on Alexnet network is proposed in this thesis.Then,from the view of optimized classification,two methods are proposed by this thesis from the angle of accuracy and training time by autoencoder: the first method is improving the existing Softmax classifier of the model,which can improve the accuracy of the model effectively;the second method is proposing a new classification model,which can reduce the training time effectively.The feasibility and effectiveness of the proposed convolutional neural network in underwater target recognition field is verified by related simulation experiments.Through theoretical research and practice in this thesis,an effective optimized model for the underwater target recognition is presented successfully.According to different application scenarios(accuracy and training time),a deeper level optimized solution is proposed by this thesis,which makes the theoretical research on underwater target recognition research has a certain meaning and value.