Region Detection Based TRA-Net for SAR Target Classification

In this paper, an attention-mechanism based target region adaptation net (TRA-Net) is designed for synthetic aperture radar (SAR) target recognition and classification. With the fact that SAR image data-acquirement has many obstacles, including poor imaging process and large scale, most of the current deep-learning methods, motivated by convolution neural network or attention mechanism, typically perform unsatisfactory on account of the problems on target area detection. The proposed TRA-Net is designed to identify the target area to the maximum extent possible, achieved by a well-trained Region Detection (RD) module. Therefore, it can provide better target characteristics for the subsequent combined target classification network. A meta-learning-based optimization algrithm is applied to the method aims to solve the limitation of the data. Through extending the data set by transformation, we achieve the same or even better classification effect in the case of limited data. Comparison tests on MSTAR and MSAR data sets have confirmed that the proposed model shows better adaptability to target region locking and target feature extraction, which obtained better target classification results.