Generalized Representation Learning of Azimuth based on Mutual Information for SAR Recognition

Automatic target recognition is of vital importance to SAR image interpretation. Nevertheless, learning-based methods still suffer from severe problems of electromagnetic scattering sensitivity and lack of data. In contrast to many methods synthesizing real-like SAR images or seeking stable features for classification, this paper focuses on the essential azimuth angle of the target, which greatly influences the SAR image due to the electromagnetic scattering mechanism. A method based on mutual information is proposed in order to achieve a balance between representative ability and generalization. The mutual information between the embedding and the azimuth is maximized, while that between the embedding and the input is minimized. Furthermore, a variational framework based on neural networks is introduced to approximate the intractable mutual information. Meanwhile, the azimuth angle is coded by a continuous vector representation to avoid discontinuity. In order to validate the generalization in open environments, a new data split from MSTAR called UTC is proposed, where unseen targets constitute the test set. Extensive experiments show that our method outperforms other discriminant models on MSTAR and exhibits superior generalization capability.