Self-Supervised Learning Combined Group and Instance Discrimination for Small Sample SAR ATR

Some Synthetic aperture radar (SAR) target images are challenging to acquire and costly to label. Self-supervised learning is available for SAR Automatic Target Recognition(SAR, ATR) applied to small samples, but the poor discrimination of SAR images and the large similarity between samples are not conducive to self-supervised learning for instance discrimination. In this paper, we proposed a self-supervised learning framework combining group instance discrimination for SAR ATR. For SAR target images, one branch is instance discrimination and the other branch is group-instance discrimination. Group instances are specifically formed by using feature clustering in a batch of instance samples and finding the centroids of groups. Then the SAR Group Instance Discriminant Loss is updated to metric the distance between the input instance eigenvectors and the centroids and converge them to the closest distance group, so that the similar eigenvectors in different clusters can be distinguished. It is experimentally concluded that this method can achieve high recognition rate with limited annotated SAR images, surpassing the performance of traditional transfer learning and self-supervised learning methods on SAR ATR tasks.