Selected Pseudo-Labelling With Class Confusion Regularization For Semi-supervised SAR Target Recognition

SAR images are highly sensitive to imaging conditions, e.g., depression angle, polarization mode and wave band. It is notable that existing enhanced structure in deep learning method does not consider domain discrepancy between training and testing data. In this paper, we employ the target domain selective pseudo-labeling to align the data distribution. To alleviate the mis-labelling issue, we introduce a novel regularization method, class confusion, the probability that neural networks misclassify the prediction between correct and ambiguous classes on target samples. Experimental results on the moving and stationary target acquisition and recognition (MSTAR) dataset can achieve state-of-the-art performance compared to other domain adaptation and model-based methods.