INTERPRETABLE DISENTANGLED ADVERSARIAL AUTO-ENCODER FOR SAR-ATR WITH SPARSE TRAINING SAMPLES

Lack of interpretability and weak generalization ability have become the major problems with data-driven intelligent Synthetic Aperture Radar-Automatic Target Recognition (SAR-ATR) technology, especially with sparse training samples. A novel insight into the typical target representation with neural networks from a causal perspective is presented in this paper. Specifically, a set of SAR images is causally modeled by intrinsic, diverse, and random attributes, which is consistent with the forward imaging process of SAR. Consequently, an Interpretable Disentangled Adversarial auto-Encoder (IDAE) is proposed based on the disentangled representations. A Symmetrically Conditional Encoding (SCE) module is established to constrain the semantic consistency of low-dimensional features. Besides, a hybrid loss function is designed for iterative training. Experiments conducted on the MSTAR dataset show that the proposed model improves both representation and generalization abilities. IDAE is able to achieve a classification accuracy of 93.1% using only 12 samples per class.