<scp>RetinexVRF</scp> : Infrared–Visible Fusion for Nighttime Low‐Light Image Enhancement

ABSTRACT In the field of low‐light image enhancement, most existing methods are optimized for scenes with weak or low illumination that still retain a certain degree of visible structural information, while relatively limited attention has been paid to images captured under extremely dark nighttime conditions. In such scenarios, images are typically characterized by a significantly reduced signal‐to‐noise ratio, degraded structural information, and missing texture details, which pose greater challenges for visual restoration. To address these issues, this paper focuses on image enhancement under extremely low illumination at night and proposes an enhancement network tailored for such conditions, termed RetinexVRF. Specifically, the Retinex theory is extended to a cross‐modal constraint framework, in which a dual‐modal feature fusion structure is designed to effectively integrate visible and infrared image information. During the enhancement stage, a vision transformer driven by fused features is introduced to perform region‐wise processing of illumination, color, and texture at specific scales. In addition, a gating mechanism is employed to selectively fuse downsampled and upsampled features, guiding the model to focus on structural restoration in key regions. Experimental results on the public LLVIP and MSRS datasets demonstrate that the proposed method improves the average peak signal‐to‐noise ratio (PSNR) by 3.47 and 4.93 dB, respectively, compared with the current state‐of‐the‐art methods. These results indicate that the proposed approach outperforms recent SOTA methods in nighttime image enhancement tasks under severely insufficient illumination.