SAR-optical image registration via diffusion model-based cross-domain translation

Image registration is one of the essential pre-processing steps in remote sensing data applications such as change detection, image fusion, and so on. Since SAR images and optical images differ in several ways but have strong complementarity, a new registration approach based on image cross-domain translation is proposed. The image cross-domain translation was realized on the foundation of the diffusion model which had a strong ability for image generation. First, the unconditional diffusion model was trained to generate SAR-like images. Then, a novel condition was introduced to the inverse process to convert the optical image into the corresponding SAR-like image via a generative network. Finally, RIFT is used to extract and match the features between the generated SAR-like image and the actual SAR image to achieve the registration between the original optical image and the actual SAR image. The experimental results show that the diffusion model-based generative network is stable in training and convenient in controlling the conditional parameters, which can better solve the problem of radiometric discrepancy between SAR images and optical images.