Edinburgh Napier University

  Synthetic Aperture Radar (SAR) target recognition is an important research direction of SAR image interpretation. In recent years, most of machine learning methods applied to SAR target recognition are supervised learning which requires a large number of labeled SAR images. However, labeling SAR images is expensive and time-consuming. We hereby propose an end-to-end semi-supervised recognition method based on an attention mechanism and bias-variance decomposition, which focuses on the unlabeled data screening and pseudo-labels assignment. Different from other learning methods, the training set in each iteration is determined by a module that we here propose, called dataset attention module (DAM). Through DAM, the contributing unlabeled data will have more possibilities to be added into the training set, while the non-contributing and hard-to-learn unlabeled data will receive less attention. During the training process, each unlabeled data will be input into the network for prediction. The pseudo-label of the unlabeled data is considered to be the most probable classification in the multiple predictions, which reduces the risk of the single prediction. We calculate the prediction bias-and-variance of all the unlabeled data and use the result as the criteria to screen the unlabeled data in DAM. In this paper, we carry out semi-supervised learning experiments under different unlabeled rates on the Moving and Stationary Target Acquisition and Recognition (MSTAR) dataset. The recognition accuracy of our method is better than several state of the art semi-supervised learning algorithms.