Optical-to-SAR image translation based on diffusion Schrödinger bridge

In the field of synthetic aperture radar (SAR) target recognition, the recognition of aircraft targets has consistently presented a significant challenge. This is primarily due to the difficulty in obtaining SAR data of aircraft. In contrast, optical images of the target are easy to obtain. Consequently, the prevailing approach has been to translate target optical images into SAR images for target recognition. With the advancement of diffusion models, the quality and diversity of generated samples have continued to improve, rendering it possible to employ diffusion models for the translation of optical images to SAR. Additionally, the Schrödinger bridge constructs a transition between two arbitrary distributions, which can be utilized in the diffusion model to achieve optimal transport with two distributions. This paper proposes an innovative solution to achieve the translation from optical to SAR images by diffusion model based on the Schrödinger bridge. Concurrently, there is an unavoidable domain shift between the generated image and the measured SAR image. In order to narrow this domain shift, a domain adaptation method is employed. The experimental results demonstrate that the data generated by diffusion can effectively assist in the recognition of three types of aircraft targets, with an accuracy rate of 71.11%.