Image Fusion for Enhanced Nighttime Visibility using Knowledge Distillation

At night, the quality of color visible images often deteriorates due to low-light conditions and the influence of artificial lighting. Enhancing the quality and effectiveness of information representation in these images can be achieved via pixel-level fusion of paired visible and infrared images. However, existing image fusion methods often fall short in rendering visually agreeable outcomes with suitable brightness and hue. In response to this, our study proposes a unified network model that combines image fusion with low-light enhancement via knowledge distillation. As a primary measure, our research advances an enhancement to the current unsupervised methodologies for improving low-light imagery during the nighttime. Following this, we develop a fusion network guided by fusion loss and color loss that brings together the enhanced colored images with their infrared counterparts. Finally, we design a lightweight student network that directly fuses nighttime infrared and color visible images. Experimental results show that the proposed method outperforms other universal image fusion methods in nighttime conditions.