Research On Underwater Sonar Image Classification With Unsupervised Domain Adaptation

Underwater sonar images are one of the main sources of underwater information and the main means of underwater target recognition today.The research of underwater sonar image classification is the premise of marine exploration and utilization.Therefore,the study of underwater sonar image classification has important significance in the fields of marine exploration and marine equipment development.Improving the classification accuracy of underwater sonar images and better adapting to the imbalance of underwater datasets are the main directions of underwater sonar image classification research.The traditional underwater sonar image is not same as the optical image.And because of the underwater noise and mechanical noise,the quality of sonar image is poor,which affects the feature extraction of underwater sonar image.Recently,deep learning,especially deep convolutional neural networks,has proven its unparalleled feature extraction capabilities.Therefore,it is necessary to apply deep convolutional techniques on underwater sonar image datasets.According to the characteristics of the underwater sonar image datasets,it is necessary to meet the conditions of unsupervised classification and the imbalanced datasets.In this way,unsupervised domain adaptation method is advisable for underwater sonar image classification.After researching related theories and methods,deep transfer learning for unsupervised classification and deep clustering networks are performed on underwater sonar image datasets.In the experiments,two generative adversarial networks are used to generate the source domain dataset,and two balanced and imbalanced original underwater sonar image datasets are also set.The comparison of results show that transfer learning can better alleviate the impact of imbalanced datasets,and the category information in unsupervised classification can improve classification accuracy.Therefore,based on the above,this paper proposes a class-conscious method for unsupervised domain adaptation(Adversarial Auto-encoder with Class-Consensus for Unsupervised Domain Adaptation,ACUDA)for unsupervised classification of underwater sonar image.By constructing a well-trained model with adversarial discriminators on the source domain,which extracts domain-invariant features through adversarial auto-encoders and convolutions,combined,the category information is embedded into the unsupervised classification process.Therefore,ACUDA achieves better performance of unsupervised classification on both balanced and imbalance sonar image datasets.The experiment evaluates the results by the confusion matrix and the comparison of results prove the effectiveness of ACUDA.In addition,the stability proof of ACUDA is provided.And the upper bound of ACUDA generalization is derived through the relationship between stability and generalization.According to the relationship between the generalization bound and the one for domain adaptation,the upper bound of the domain adaptive generalization of ACUDA is also given in theories with optimization goals.Finally,it is proved that the proposed ACUDA has a clear upper bound of generalization for domain adaptation.