Underwater Target Recognition Based On Generative Adversarial Networks

To continuously deepen the exploration and cognition of the ocean,the rapid development of underwater target recognition is particularly important.How to accurately and quickly identify targets is the key to the process of exploring marine resources and national defense.In recent years,the popularity of deep learning has continued to rise and the scope of application has continued to expand,and it has been gradually applied to underwater target recognition.However,samples of underwater targets are often difficult to obtain,so in order to solve the problem of poor recognition performance of deep learning under small sample conditions,this paper uses a Generative Adversarial Networks to solve it.The main research contents and methods are as follows:The related principles of Convolutional Neural Networks are studied.In order to discuss the influence of the number of sample sets on the recognition performance of Convolutional Neural Networks,a comparative experiment on the influence of the number of sample sets on the recognition rate of Convolutional Neural Networks is carried out,and the experimental results are analyzed.Considering that the Generative Adversarial Networks has the function of expanding the sample set,two types of Generative Adversarial Networks models are studied: deep convolutional Generative Adversarial Networks and ACGAN.In order to compare the performance of the generated samples of the two generational confrontation models,a comparative experiment on the performance of the generated samples based on the two network models was carried out,and the analysis and performance of the experimental results were carried out from the loss curve of the training process,the visualization during the training process,and the evaluation criteria of the generated samples.In the evaluation,it is concluded that the quality of ACGAN generated samples is better.Carried out the network structure comparison experiment and parameter comparison experiment in ACGAN,and obtained more suitable parameters and structure through a series of comparison experiments.Then the process experiment of ACGAN generating samples was carried out,and the experimental results were analyzed.Under the condition of the small sample,the comparison experiment of the recognition rate of the generation module and the discrimination module in ACGAN when the generation module and the discrimination module work together and when the discrimination module works alone,it can be concluded that the recognition performance of ACGAN is better than that of the discrimination module alone under the small sample condition Recognition performance at work.