MAFNet: Deep Merged Autofocusing Network for SAR Recognition Under Defocused and Noised Dataset

Airborne synthetic aperture radar (SAR) imagery is susceptible to non-systematic motion errors, which will definitely defocus the target image and degrade the recognition accordingly. Representations learned from conventional convolutional neural networks (CNNs) do not emphasize much about the intrinsic features of the data distribution, which collapses the purposefulness of feature extraction. To this end, a deep Merged Auto-Focusing Network (MAFNet) is proposed for adaptively removing the defocusing effect from the input SAR data used for target recognition. Specifically, we propose an end-to-end architecture consisting of one focusing module and one recognition module. The focusing module contains a U-shaped sub-network based on the cross convolution to ensure the phase history coherence of the extracted features, and a deep unfolding method is devised to improve the mathematical generalization of focusing and sparse features so as to enhance the purposefulness of feature extraction. In particular, a compact surrogate function is designed for the non-convex feature quantization problem in the focusing module, which leads to closed-form solutions. The recognition module simply consists of a classifier. By building a multivariate loss function consisting of a cross-entropy loss function and a novel focusing loss function, MAFNet achieves accurate target recognition even when the input data is contaminated by non-systematic phase errors and additive noises. MSTAR data set is utilized to validate the effectiveness of MAFNet, and comparisons with conventional algorithms are performed to demonstrate the superiority of the proposed network.