Learning to Generate SAR Images With Adversarial Autoencoder

Deep learning-based synthetic aperture radar (SAR) target recognition often suffers from sparsely distributed training samples and rapid angular variations due to scattering scintillation. Thus, data-driven SAR is considered a typical few-shot learning (FSL) task. This article first reviews the key issues of FSL provides definition A novel adversarial autoencoder (AAE) then proposed as an representation generation network. It consists generator network that decodes knowledge images discriminator not only learns discriminate “fake” generated real ones but also encodes input image back knowledge. The employs progressively expanding convolution layers corresponding layer-by-layer strategy. uses two cyclic loss functions enforce consistency between inputs outputs. Moreover, rotated cropping introduced mechanism address challenge representing orientation. moving stationary Target (MSTAR) 7-target dataset used evaluate AAE’s performance, results demonstrate its ability generate with aspect diversity. Using 90 at least 25° orientation interval, trained AAE able remaining 1748 other angles unprecedented level fidelity. it can be for data augmentation in tasks. Our experimental show could boost test accuracy by 5.77%.