(A) study on synthetic-to-measured SAR target image translation based on transformers

Abstract

Automatic target recognition (ATR) in synthetic aperture radar (SAR) images has been actively studied in recent years. However, it is time-consuming and costly to collect large amounts of labeled SAR data with complex processes, and thus training high-performance SAR ATR network based on a data-driven approach is challenging. In order to solve the problem of insufficient SAR data, a method using synthetic SAR data was proposed. However, despite sophisticated modeling, there is a domain gap between synthetic SAR data and measured SAR data, which indicates that the target classification model trained with synthetic SAR data tends not to be generalized to measured SAR data. Therefore, in this thesis, we propose a transformer-based SAR target image translation network that can effectively reduce the domain gap between synthetic and measured SAR data by considering the characteristics of synthetic SAR. A proposed model which consists of three sub-networks learns and combines mapping in image space and mapping in feature space to translate simulated SAR images into measured SAR images. In addition, we propose five scenarios that can validate the performance of the image translation model using a limited SAR dataset. Through these scenarios, we verified that our model effectively translates synthetic SAR data to measured SAR data, and also demonstrates well generalization ability on unknown synthetic SAR images such as unknown target classes and azimuth angles.