SAR Target Classification Based On Generative Adversarial Network And Capsule Network

Synthetic Aperture Radar(SAR)is a system capable of active observation of targets at all time,all weather and at a distance.It has been widely used in military and civilian fields.SAR Automatic Target Recognition(SAR-ATR)is one of the research hotspots in SAR field.It can automatically,quickly and accurately identify targets.In recent years,deep learning theory has been widely used in SAR-ATR,and has achieved remarkable results.Convolutional Neural Network(CNN)is a typical deep learning network.It can obtain high target recognition rate by extracting different levels of target features.However,SAR target classification based on convolutional neural networks requires a large number of training samples,otherwise it is prone to overfitting problems.In addition,CNN cannot solve the problem of misclassification caused by the change of target posture(such as translation,rotation,scaling,etc.).In order to solve these problems,this paper combines the advantages of generat adversarial networks and capsule networks,and proposes two SAR image target recognition methods.The specific contents are as follows:1)A SAR target classification method based on improved Convolutional Neural Network(ICNN)and improved Generative Adversarial Network(IGAN)is proposed.In other words,the IGAN is pre-trained unsupervised with training samples,then the ICNN is initialized with the trained IGAN discriminator parameters,then ICNN is fine-tuned with training samples,and finally the trained ICNN is used to classify the test samples.The experimental results of MSTAR tenth class show that the proposed method can not only achieve a recognition rate of up to 96.37% when the number of training samples is reduced to 30% of the original sample,but also has stronger anti-noise performance than the method using ICNN directly.2)A SAR target classification method based on an improved Capsule Network(ICapsNet)was proposed.ICapsNet consists of two capsule networks,each of which use convolution kernel of different size to extract different level features of target,and introduce an attention module to enhance the features of different level.Finally,the digital capsule output of the twochannel network is summed to further increase the probability of the category to which the target belongs.The results of MSTAR classification experiment show that the proposed method can achieve a high recognition rate under the condition of a small number of training samples and a change in the attitude of test samples.