DMWFuse: A degradation-adaptive unified framework for RGB-IR image fusion

In recent years, object detection using visible (RGB) and infrared (IR) images has garnered widespread attention as a promising solution for achieving robust detection in real-world application scenarios. However, real-world imaging is frequently plagued by degradation issues-such as noise, blur, and haze in visible light imaging, as well as stripe noise in infrared imaging-which significantly degrade model performance. To address the research challenges in infrared-visible image fusion under degraded scenarios, this paper proposes a unified framework unifying multi-modal image restoration and fusion, adopting a two-stage training strategy. Its core innovations include a Task-Aware Temporal Segmented Mixture of Experts Diffusion Module (TDMoE), which optimizes expert selection through a gating mechanism to adaptively learn various degradation types during the restoration stage and aggregate high-value multi-modal information during the fusion stage; a Wavelet Domain Fusion Module (WFM) enabling differential fusion of structures and details; and a Wavelet-Enhanced Transformer Block (WETB) that strengthens spatial-frequency domain information interaction and suppresses degradation-induced noise. The proposed network reduces task interference through two-stage collaborative learning, and extensive experiments on multiple RGB-IR datasets validate that our method achieves superior performance and robust generalization for downstream tasks. Code is available at https://github.com/M2FLab/DMWFuse.