Physically Driven Source-Aware Decoupling and Degradation-Aware Diffusion for Infrared Image Enhancement

Infrared imaging for remote sensing follows a complex physical pipeline that introduces multiple coupled degradations, making its robust enhancement fundamentally distinct from conventional RGB image restoration. Most existing infrared enhancement methods target a single degradation type, and their extension to joint multidegradation handling often suffers from severe error accumulation and mutual interference. To address this challenge, we propose source-aware degradation decoupling diffusion framework for infrared image enhancement (SDDiff-IR), a novel physics-driven end-to-end infrared enhancement pipeline for joint multisource degradation removal. Motivated by the infrared imaging physical process, SDDiff-IR first uses the physically inspired decomposition network to decompose degraded infrared images into two physically interpretable components: a radiation map and a detail map, corresponding to scene-to-detector and detector-to-signal degradations, respectively. The radiation map mainly captures low-frequency radiation bias caused by scene radiation, atmospheric effects, and the optical system. To correct the radiation bias, we design a generative diffusion model to reconstruct the underlying radiation distribution. On the contrary, the detail map characterizes high-frequency degradations originating from the infrared focal plane array (FPA) and readout circuitry, including stripe and Gaussian noise. We introduce a degradation-aware guided diffusion model that jointly learns degradation synthesis and restoration, enabling effective guidance of the diffusion process for modeling and removing multiple noise sources. Extensive experiments on public real-world infrared datasets demonstrate that SDDiff-IR consistently outperforms state-of-the-art infrared enhancement methods in both quantitative and qualitative evaluations.