Binary Diffusion Method for Image Fusion with Despeckling

Remote sensing image fusion is the integration of multimodal information to facilitate the comprehensive acquisition of target characteristics for image interpretation and analysis. Recently, the diffusion model and its continuous form have shown good performance in various image processing fields due to their advanced generative capability, rigorous interpretability and easy generalization in the pre-training framework. This paper focuses on the fusion of optical and SAR images. However, due to its operating principle, SAR is inevitably affected by speckle noise, which limits the performance of image fusion and affects the subsequent interpretation of information from the fused image. To improve this issue, an image fusion algorithm that can suppress speckle noise from SAR images is proposed. First, to get the true distribution of the SAR image without speckles, which is equivalent to the task of despeckling, the variational optimization model is established by variational inference. Then, with the fidelity term of variational optimization model served by the speckling distribution and the regularization term served by the matching score function during the denoising of the diffusion model, first-order stochastic optimization is used to solve optimization problem. Thirdly, with similar method, the optimization model of image fusion in timestep t is established. Finally, the binary diffusion model of image fusion with despeckling is modeled by binary optimization in every timestep, which achieves both image fusion and despeckling. The experimental results show the proposed method has obvious visual anti-noise effect and significant improvement in image quality metrics.