Deep Transfer Learning Based on Generative Adversarial Networks For SAR Target Recognition with label limitation

Abstract

Driven by the large-scale labeled dataset, convolutional neural networks(CNNs) have achieved great progress in optical image classification. Owing to the robust ability of image feature extracting, deep CNNs are also suitable for Synthetic Aperture Radar automatic target recognition(SAR-ATR) tasks. However, the labeled SAR image sets are always too small, which often lead to a severe over-fitting while training a deep CNNs. In this paper, we propose a novel transfer learning method with generative adversarial networks(GANs) to solve the above problem. Instead of training a deep neural network with the insufficient labeled dataset, we train a GAN using a large numbers of unlabeled SAR images to learn generic features. Then, the pre-trained layers are reused to transfer the generic knowledge to specific recognition tasks through the fine-tuning technique. The experimental results demonstrate that the transfer learning method has a particular performance improvement in SAR-ATR tasks and has less performance degradation when the dataset becomes even smaller. The experimental results prove the feasibility of using GANs for transfer learning, which is important for further research to improve the performance of SAR-ATR with label-limited datasets.