Sar Automatic Target Recognition Based On Complex-Valued Fully Convolutional Neural Network

Synthetic Aperture Radar(SAR)is a kind of high resolution imaging radar.It is not affected by illumination and climate conditions.It can monitor ground targets all day and all weather.It plays an important role in civil and military fields such as resource exploration,disaster monitoring and target tracking.SAR Automatic Target Recognition(SAR-ATR)has always been the focus of research in various countries as an effective means of SAR image interpretation.In recent years,with the rapid development of deep learning,Convolutional Neural Network(CNN)has been widely used in SAR target recognition.In this paper,Fully Convolutional Neural Network(FCNN)is studied in depth on the basis of CNN,and the framework of FCNN is extended to Complex-valued Fully Convolutional Neural Network(CV-FCNN).In order to obtain higher target recognition rate with a small number of samples,the SAR automatic target recognition method combining improved Convolutional Auto-Encoder(ICAE)and FCNN is further studied,and further extended to the CV-FCNN framework.The main work of this paper is as follows:(1)For MSTAR complex data containing amplitude and phase information,a framework of CV-FCNN is proposed.On the basis of CV-FCNN,a convolution layer with 1×1 convolution kernels is added as an intermediate network to improve the nonlinear operation of local receptive field operation in the complex convolutional layer.The experimental results show that the average correct recognition rate of MSTAR target recognition reaches 99.71%.(2)Aiming at the problem of scarcity of labeled data samples in SAR target recognition,a method of SAR target recognition based on FCNN and ICAE is proposed,which is further extended to the CV-FCNN framework.The experimental results show that,without expanding the training samples,the recognition rate of ten kinds of targets using MSTAR amplitude information reaches 98.14%,and the average correct recognition rate of ten kinds of targets using complex MSTAR containing amplitude and phase information reaches 99.18%.In addition,the experimental results also show that the method has a certain anti-noise ability.