Consistency regularization teacher–student semi-supervised learning method for target recognition in SAR images

Synthetic aperture radar automatic target recognition (SAR-ATR) is a hotspot in the field of remote sensing, which has been widely used in disaster monitoring, environmental monitoring, resource exploration, and crop yield estimates. In recent years, deep convolutional neural networks (DCNNs) have achieved promising performance among a variety of supervised classification methods under the condition of sufficiently labeled samples. However, it is expensive and time-consuming to collect large amounts of labeled samples suitable for DCNNs in SAR domains. To reduce the dependence of SAR-ATR on labeled samples, in this work, a consistency regularization teacher–student semi-supervised (CRTS) method for SAR-ATR is proposed, in which consistency regularization is applied to analyze and divide unlabeled samples and the teacher–student structure is introduced to generate pseudo-labels for the divided unlabeled samples. Firstly, by using consistent pseudo-label prediction, unlabeled samples are divided into consistent unlabeled samples and confident unlabeled samples. Then, in order to improve the quality of pseudo-labeled labels, the student model is used to generate a pseudo-label for consistent unlabeled samples, and the teacher model which is ensembled by the other two student models labels the confident unlabeled samples. Finally, these pseudo-label unlabeled samples are mixed with the labeled samples and trained together to improve recognition performance. Experiments are conducted on the MSTAR dataset, and the results demonstrate the effectiveness of the proposed method. As compared with several state-of-the-art methods, the recognition accuracy shows the superiority of the proposed method, especially when the training dataset is limited.