Retinex-based Low-Light Image Enhancement

Abstract

Image enhancement in low-light conditions is not solely a matter of adjusting the illumination. Due to the limited capabilities of capturing devices, there are also issues of hidden noise bursts and color distortion in dark environments. The Retinex model, which decouples observed image space into different subspaces for image enhancement, has been extensively validated as effective. Inspired by this model, this paper proposes a low-light image enhancement method based on Retinex model decomposition, guided by deep neural networks.The proposed approach decomposes the image into the illumination map responsible for light adjustment and the reflectance map responsible for removing degradation. A light enhancement network is trained using image pairs captured with different exposures to achieve user-friendly adjustable light intensity enhancement. To alleviate the issue of color distortion, a reflectance restoration network incorporates illumination guidance information. Additionally, a diffusion model is employed as a noise detector to eliminate unnecessary noise and other distortions in the brightened low-light images by fitting the conditional distribution between different images. Experimental results demonstrate that the proposed method outperforms in terms of color preservation in low-light images with varying light intensities. It effectively removes noise while preserving image details.