SARCUT: Contrastive learning for optical-SAR image translation with self-attention and relativistic discrimination

Abstract

Focusing on how to obtain high-quality and sufficient synthetic aperture radar (SAR) data in deep learning, this paper proposed a new mothed named SARCUT (Self-Attention Relativistic Contrastive Learning for Unpaired Image-to-Image Translation) to translate optical images into SAR images. In order to improve the coordination of generated images and stabilize the training process, we constructed a generator with the self-attention mechanism and spectral normalization operation. Meanwhile, relativistic discrimination adversarial loss function was designed to accelerate the model convergence and improved the authenticity of the generated images. Experiments on open datasets with 6 image quantitative evaluation metrics showed our model can learn the deeper internal relations and main features between multiple source images. Compared with the classical methods, SARCUT has more advantages in establishing the real image domain mapping, both the quality and authenticity of the generated image are significantly improved.