DBD-CR: Dual-Branch Diffusion Residual Reconstruction for Cloud Removal in Optical Remote Sensing Images

Optical remote sensing image cloud removal is a crucial area of research within remote sensing technology. Its primary goal is to recover surface information from images obscured by clouds. Although existing generative models have achieved significant success in cloud-free image reconstruction, providing important support for downstream tasks, challenges remain in capturing and reconstructing complex detailed features. To address these issues, this article presents a dual-branch diffusion residual reconstruction framework for optical remote sensing image cloud removal, called the DBD-CR framework. Specifically, in the initial reconstruction stage, we utilize a diffusion model to obtain the initial reconstructed image from the cloudy input. Next, in the discrepancy correction stage, we calculate the residual texture map between the cloudy image and the initial reconstructed image to capture the details of the surface features that have not been restored. In addition, we integrate a frequency domain enhancement interaction attention module and an adaptive mixed attention module (AMAM) into the denoising diffusion model. These components capture crucial details in the frequency domain and adaptively select features that combine downsampling and upsampling. By adopting more comprehensive evaluation metrics and conducting extensive experiments on three public remote sensing cloud removal datasets, the DBD-CR framework achieves a maximum reduction of 18.081 in the Fréchet inception distance (FID) metric and a maximum improvement of 0.902 in the peak signal-to-noise ratio (PSNR) metric, further demonstrating its advanced performance for optical remote sensing image cloud removal tasks.