SAR Target Recognition Based on Decoupling and Reconstruction Learning Using Complex-Valued Convolutional Neural Networks

Recent years, the deep convolutional neural networks (CNNs) based SAR target recognition methods are springing up. However, existing CNN-based SAR target recognition methods are limited by down-sampling caused information loss and feature redundancy, leading to learn shallow data correlations and bringing the performance collapse in practical applications of SAR target recognition. In this paper, a complex-valued decoupling and reconstruction learning (DRL-Net) based method is proposed to improve the SAR target recognition performance. DRL-Net learns the decoupling representation from the complex-valued images to make sure the learned latent feature is sparse and informative. The learned latent feature is then fed to reconstruction learning to keep the information completeness. Finally, the discriminative information is powered into the latent features to gain the semantic for SAR target recognition. Compared with baseline method, DRL-Net gains higher accuracy and faster convergence with the same number of samples. Additionally, DRL-Net exhibits a good potential for SAR target recognition with limited samples.