Selecting pseudo supervision for unsupervised domain adaptive SAR target classification

Abstract In recent years, deep learning has brought significant progress for the problem of synthetic aperture radar (SAR) target classification. However, SAR image characteristics are highly sensitive to change imaging conditions. The inconsistency parameters (especially depression angle) leads distribution shift between training and test data severely deteriorates classification performance. To address this problem, in paper we propose an unsupervised domain adaptation method based on selective pseudo-labelling Our directly trains a model using from by generating pseudo-labels domain. key idea is iteratively select valuable samples optimize classifier. each iteration, breaking ties (BT) criterion adopted best with highest scores relative confidence. Besides, avoid error accumulation iterative process, class confusion regularization used improve accuracy pseudo-labelling. compared state-of-the-art methods, including supervised over moving stationary acquisition recognition (MSTAR) dataset. experimental results demonstrate that proposed can achieve better performance, especially when difference angles source images large. our also shows its superiority under limited-sample