Similarity to Availability: Synthetic Data Assisted SAR Target Recognition via Global Feature Compensation

Due to the scarcity of synthetic aperture radar (SAR) images, the SAR target recognition can be regarded as few shot problem, one of the solutions is to generate synthetic SAR images by computer simulation. However, the synthetic data are similar but unavailable, because there is a domain gap and distribution differences between synthetic and measured SAR images, causing inability to directly apply the synthetic data for SAR target recognition. To solve this problem, the global feature compensation network based on the generative adversarial network is proposed, translating the synthetic data from similarity to availability and assisting SAR target recognition. First, the background feature compensation module is proposed, to compensate for the low-level texture features via the feature skip connection fusion, and associate the entire image features via Swin transformer during the feature extraction stage. Then, a window multihead mix-attention module is proposed to improve the network's learning ability of the measured SAR target features, by assisted with the synthetic SAR data. Experiments on the Synthetic and Measured Paired Labeled Experiment dataset show that the proposed method can well learn the potential features of the SAR target and translate the synthetic SAR images closer to the measured SAR images. According to the experimental results, the proposed method outperforms the State-of-the-Art translation methods and SAR automatic target recognition algorithms for limited training samples.