Research On The Classification Method Of Side-scan Sonar Image Based On Transfer Learning

71% of the earth is covered by the ocean,which is closely related to human beings.In life,people need to explore the seafloor topography and various resources needed;in national defense,people need to detect some military targets in the sea to make defense.Therefore,it is very important to detect and recognize the ocean target.Because the image of optical image in water is fuzzy,easy to be interfered by light refraction and reflection,and the imaging distance is limited,the detection of seabed target is mainly based on side scan sonar image.At present,most of the existing image classification methods are suitable for optical images,among which the traditional methods are sift,Harrs and other feature extraction algorithms combined with neural network,Bayesian and SVM classification methods to complete the image classification.These methods require people to have a certain understanding of the image,and people need to rely on experience and specific sonar image to extract useful features for classification Therefore,the effect is not good.In order to solve the above problems,this paper uses a new method and improves it to improve the classification performance of side scan sonar image.Firstly,the convolution neural network method is used to automatically extract the features of the side scan sonar image.Because the data of side scan sonar image is too few to meet the data requirements of convolution neural network,this paper expands the image based on the existing side scan sonar image.In the experiment,different types of convolutional neural networks are trained based on the original data set and the expanded data set,and the classification accuracy of both is much higher than that of the traditional method;at the same time,the classification accuracy of the expanded data set is higher than that of the original data set,which verifies the validity of the data expansion.Secondly,in view of the problem that the number of side scan sonar images is too small,a multi task learning method based on convolution neural network is used to assist the classification of side scan sonar images by using the data in the source domain;in view of the difference between the bottom features in the source domain and the target domain,based on the original method,the bottom feature extractor is separated,and an improved multi task learning method is proposed.In the experiment,the network proposed by the two methods is trained on the side scan sonar data set.The experiment shows that the multi-task learning method is better than the convolution neural network method,and the improved multi-task learning method is better than the general multi-task learning method in classification effect.Finally,in order to solve the problem that the domain difference between the source domain and the target domain is not considered in the multi task learning,the transfer learning method based on depth adaptation is used.The depth adaptation method reduces the gap between the fields in the training process,so that the knowledge in the source field can be applied to the side scan sonar image.Aiming at the problem that the depth adaptation network does not use the label of the side scan sonar image,the depth adaptation method is improved from two aspects.Firstly,multi-core MMD is used to calculate the distance between the same kind of samples in different fields instead of the original method;secondly,a method combining multi-task learning and transfer learning is proposed,which adds the task of opposite scan sonar image classification based on the transfer learning network.In the migration learning office-31 data set and the side scan sonar image data set,the effects of various methods are compared,and the effectiveness of the migration learning method and the improved method is proved.