A Dual-Stage Residual Diffusion Model With Perceptual Decoding for Remote Sensing Image Dehazing
Hao Zhou, Yalun Wang, Qian Zhang, Tao Tao, Wenqi Ren
Abstract
Atmospheric pollutants, such as haze, severely affect the quality of remote sensing images, leading to blurred details and impairing their effectiveness in applications like environmental monitoring and agricultural resource management. In recent years, diffusion models have attracted widespread attention due to their powerful generative capabilities. However, striking a balance between their expensive training costs and actual recovery effectiveness has become a major challenge. To address this key challenge, we propose a perceptual decoding dual-stage residual diffusion model (DS-RDMPD) for remote sensing image dehazing. The core innovation of our work lies in a dual-stage coarse-to-fine architecture that integrates the traditional U-Net with a diffusion model, enabling efficient adaptation of diffusion-based restoration to the dehazing task. This design not only achieves strong performance but also demonstrates remarkable generalization across various image restoration scenarios. In the first stage, we use the Multi-channel Efficient Selective Synthesis U-Net (MCESS-UNet) to pre-process the remote sensing haze images. This architecture performs initial dehazing and feature extraction through a multi-scale channel attention (MC) block, and then performs enhanced spatial feature aggregation through an Efficient Selective Synthesis (ESS) block. The preprocessed image is then used as the conditional input of the Residual Diffusion Model with Perceptual Decoding, where the perceptual decoder improves the generation quality by further decoupling the condition to refine the residual estimate. Extensive experiments on multiple datasets show that DS-RDMPD can achieve satisfactory results with only 300,000 iterations and about five sampling steps. It has achieved satisfactory results in both qualitative and quantitative experiments, and also exhibits strong performance in rain removal and deblurring tasks, demonstrating the excellent generalization ability of the model. The code is available at https://github.com/Aaronwangz/DS-RDMPD.