Transfer Learning With Deep Convolutional Neural Network For SAR Target Classification

With the widespread use of synthetic aperture radar(SAR)system in military and civil fields,the scale of SAR image data has rapidly expanded,and the target classification requirements for SAR images in different application scenarios have also increased.Traditional target classification algorithms need to design feature extractors manually according to image characteristics after analyzing data sets.The design process is complicated and cumbersome and has serious dependence on professional knowledge,which is difficult to meet the actual needs.Therefore,the deep learning method has been introduced into the SAR image target classification field.However,due to the small scale of the existing tagged SAR image dataset and the difference in image features between the SAR image and the optical image,it is often difficult to obtain the ideal effect by directly applying the optical image convolutional neural network model to the SAR image.To solve the above problems,this paper proposes a SAR image target classification method based on convolution neural network transfer learning.The main work of the thesis is as follows:1 ?The small sample size problem is a difficulty in the practical application of convolutional neural network algorithm,which is especially prominent in the SAR image classification task with small data sets.In order to apply the convolution neural network algorithm to SAR image classification task better,in this paper,we propose a SAR image target classification method based on convolution neural network for military and civilian vehicle classification tasks.From the perspective of model structure,this method compresses the model parameter quantity and adjusts the network convolution structure to solve the problem of small sample size and the difference between optical image and SAR image.And based on the existing optical convolution neural network model,a convolution neural network model suitable for SAR image is proposed.The experimental results based on MSTAR and GOTCHA datasets verify the accuracy of the model in SAR image classification tasks.2?Compared to SAR images,optical image data is more widely available and easier to acquire.Another way to solve the problem of small sample size is using pre-training model,which is fully trained on large-scale optical data sets,to assist the training of SAR image classification model through parameter-transfer.However,the classification model designed for SAR images is very different from the existing optical convolutional neural network model.Traditional transfer learning methods can not realize the parameter-transfer between networks with different structures.Therefore,in this paper,we propose a SAR target classification method based on parameter-transfer between networks with different structures.This method combines the source domain model and the target domain model as two branches to form a two-stream network.In training phase,we constrain the output of the two branches to be approximate or identical when the input data is consistent,thereby achieving the knowledge transfer.The experimental results based on MSTAR dataset verify the effectiveness of this method.