PDDM: Prior-Guided Dual-Branch Diffusion Model for Pansharpening

Abstract

Pansharpening is to fuse a panchromatic (PAN) image with a multispectral (MS) image to obtain a high-spatial-resolution MS (HRMS) image. Although the denoising diffusion probabilistic model can generate high-quality image details, its inherent stochasticity can lead to spectral and spatial distortions in the pansharpening task, and the adding noise method for fixed-size images can weaken the generalization of the model at different scales. To address these issues, a novel pansharpening method based on prior-guided dual-branch diffusion model (PDDM) is proposed. First, a dual-branch diffusion model for different information flows from MS and PAN images is constructed to achieve the spatial and spectral fidelity, which is developed by a collaborative and adversarial learning strategy. Then, to guide detail recovery and reduce the uncertainty of the generated detail information, two pregeneration modules based on different prior information are designed for pixel-to-pixel reconstruction. Finally, a focus module is constructed to fuse the features from the dual-branch and improve the generalization of the proposed PDDM. Extensive experiments on multiple satellite datasets demonstrate that the proposed PDDM has superior performance compared to state-of-the-art methods.

Keywords

• Diffusion model
• dual-branch
• pansharpening
• pregeneration module