SAR Target Recognition Under Small Sample Via Data Generation

Synthetic Aperture Radar(SAR)is widely used in both military and civilian fields because of its all-weather and all-weather characteristics.With its excellent feature extraction capability,deep learning is widely used in SAR target recognition.However,data-driven deep learning method requires a large amount of data as support.Because of the change of operation mode in modern war,it is very important to obtain the information of the type and model of military target.However,as a high-value strategic target,SAR image data has the problems of confidentiality and expensive resources,resulting in the lack of effective and marked measured data in reality.The small sample problem of SAR target recognition restricts the performance of SAR image classification,which has important practical significance for the research of expanding the data set by means of data generation to improve the recognition ability of the model.Based on the generated adversarial network model,this thesis studies the SAR target recognition method based on the generated adversarial network in small samples.The specific research content and main work include the following aspects:(1)Aiming at the problem that the control and training of a single attribute are unstable in the current generation process of SAR target generating countermeasure network,a SAR target sample generation method based on multi-attribute directional control is proposed in this thesis.This method maximizes the mutual information between the potential codes and these attributes in the training process by Info GAN to establish a strong correlation between them and decouple the desired SAR image attributes to achieve the generation of directional control samples.By introducing category labels and corresponding classifiers to guide training,the generated images are conducive to supervised ATR models.In order to effectively measure the fitting effect between the real sample and the generated sample distribution,Wasserstein distance measure was used to describe the distance between them.In order to solve the problems of gradient disappearing and unstable training,gradient penalty and spectrum normalization were introduced to constrain the network training parameters under the condition of multi-attribute trait sample generation.The Pytorch deep learning framework was used to complete the experiment,complete the feature decoupling work of SAR image,realize the directional control of multiple attributes of azimuth Angle,scattered area and scattered intensity,and realize the high-quality generation of SAR target samples under the condition of small samples.(2)Aiming at the negative impact of generation quality instability on the recognition model and the effective utilization of generated samples,this thesis proposes a network interpretable GAN generated sample screening method to screen SAR images with specific types of typical features.This method compares SHAP values of the same feature between different generated samples by SHAP visualization method,selects generated samples corresponding to category labels and features,and screens out highly available samples for ATR model.Through qualitative and quantitative analysis,the effectiveness of the generation method and screening method is demonstrated.The SAR target samples after screening are used in the ATR model,which effectively improves the performance of the recognition model under the condition of small samples.