SAR object classification using the DAE with a modified triplet restriction

Abstract

Although deep learning methods have made great progress in synthetic aperture radar (SAR)-based remote sensing, lack of training data has often been the major obstacle while they are adopted for SAR automatic target recognition. In this study, a new deep network in the form of a restricted three-branch denoising auto-encoder (DAE) is proposed to take the full advantage of limited training samples. In this model, a modified triplet restriction that combines the semi-hard triplet loss with the intra-class distance penalty is devised to learn discriminative features with a small intra-class divergence and a large inter-class divergence. Besides, the reconstruction distortion is measured between the model outputs and the images filtered by the improved Lee Sigma filter rather than the original inputs to suppress clutter in the background. Furthermore, a batch-based triplet loss, which calculates the modified triplet loss in a batch-based manner, is proposed to tackle the difficulties in implementation and reduce its computation complexity. The simplified version of the three-branch Triplet-DAE is subsequently devised as a one-branch DAE restricted by the batch-based triplet loss. Experimental results with the MSTAR data demonstrate the effectiveness of the proposed method on real SAR images.