Research On Underwater Target Recognition Method Based On Transfer Learning

With the advance of science technology and the need of modernization of navy,underwater target recognition has become one of the hot spots in the world.Underwater target recognition and its related technology are very important to enhance the intelligence level of naval equipment,also they have significant research value and significance.In naval warfare,in order to avoid being discovered and attacked by hostile,it was impossible to track the target and collect information at short range for a long time,which lead to less data of the related targets.In the case of the existing underwater target recognition algorithm,the prediction precision of the classifier was not satisfactory,while the transfer learning can be used for knowledge learning across the domain,effectively utiliza the large amount of data in the related fields,and solve the problem of less data of training samples.Therefore,this article will introduce the transfer learning into the underwater target recognition,in the case of limited underwater target data,a large amount of data information in the source domain is migrated to the target domain,and the data quantity used to train the classifier is enlarged,which can effectively improve the performance of the classifier.In this thesis,the transfer learning technology and underwater target recognition technology are combined to solve the problem of underwater target recognition,which is often encountered in two scenarios,proposed a specific solution.For the depth unsupervised metric domain adaptation algorithm,the measurement distribution distance difference adopted the maximum mean difference,only the first order statistic of the data distribution was considered,and the distance metric function was sensitive to the selection of kernel function when the data distribution difference is large.In this thesis,a depth multi kernel adaptive algorithm for underwater target recognition is introduced,which is based on divergence deviation metric and multi-core technology.The improved algorithm transforms the sound signal collected by sonar into the sound spectrum,which can be used to aid the training classifier of the target domain with a large amount of sample data from the source domain,and effectively improve the classification accuracy.The improved algorithm effectively solves the problems that the data distribution is not considered sufficiently and the single kernel function is not definite the optimal bandwidth of the Gaussian kernel function.For the depth semi-supervised domain adaptation algorithm,only the single layer full connection layer was suitable,the problems of distinguishing information from hidden layer was not fully utilized,and the distance between Euclidean distance metric points couldn't be overcome,and a new depth semi-supervised domain adaptation algorithm for underwater target recognition was proposed.The problem of low precision of the existing underwater target recognition method is solved by changing the adaptive monolayer full join layer to fit the multilayer full connection layer and changing the Euclidean distance to the standard Euclidean distance in the target domain with the lack of the label sample data.In this thesis,by using the underwater target data set,the current mainstream Transfer learning algorithm is applied to the underwater target recognition task,the algorithm is compared with the experimental results of the underwater target recognition algorithm in two scenes of underwater target identification.It is verified that the two algorithms presented in this thesis have obtained the best recognition effect respectively in their respective scenes.