A Novel Teacher–Student Framework With Degradation Model for Infrared–Visible Image Fusion

The fusion performance of infrared (IR) and visible (VIS) images depends on the quality of the source images, which are often affected by some factors in real-world scenarios, such as environmental changes, hardware limitations, and image compression. The influence of these factors can be minimized by training a neural network capable of generating high-quality (HQ) fused images from low-quality (LQ) source images. However, in real-world conditions, it is challenging to acquire paired LQ source images and their corresponding HQ fused images for network training. To address this issue, we propose a novel teacher–student framework with a degradation model for IR–VIS image fusion, namely, TSDM-Fusion. In this framework, the teacher network is utilized to generate HQ RGB fused images, while the degradation model is employed to generate LQ IR and VIS images. Subsequently, the obtained images are used to train the student network, enabling it to learn the mapping from LQ IR and RGB VIS images to HQ RGB fused images. The design for the degradation model is the most important part of the framework, which simulates degradation factors in the imaging processes, including brightness, contrast, blur, noise, and JPEG compression. Experiments on multiple public image fusion datasets and
$\mathrm {M^{3}}$
FD detection datasets demonstrate that our method can not only enhance the visual fusion effects but also improve the detection mAP. The code and pretrained models are available at
https://github.com/bearxwm/TSDM-Fusion
.