Enhanced Infrared-Visible Image Fusion Method with an ISP Based Degradation Model

The performance of infrared (IR) and visible (VIS) image fusion tasks depends on the quality of source images, which are usually impacted by various degradation factors in real-world scenarios, leading to poor quality and visual effects of the fused images. By learning the mapping from low-quality (LQ) source images to high-quality (HQ) fused images, neural networks can remove the impacts of degradation factors, generating HQ fused images. However, the lack of paired LQ source images and corresponding HQ fused images poses a significant challenge in training a fusion network. Considering the effectiveness of the image processing pipeline (ISP) in processing IR and VIS images, reversing specific steps of ISP can accurately simulate the image degradation process. Building upon this research, we propose a novel ISP-based degradation model that reverses several ISP steps while incorporating various degradation factors, thereby generating LQ IR and VIS images from source images to train the fusion network. Experimental results on multiple image fusion datasets demonstrate that our method significantly enhances the quality and visual effects of fused images.