Low-light Image Enhancement via Dual Reflectance Estimation

Improving the quality of low-light images is a fundamental task with vast applications in computer vision. Retinex-based methods which decompose the images into reflectance and illumination components have been actively studied over the past years. In this paper, we propose a Retinex-based method with dual reflectance estimation. To be precise, we start with a simple reflectance estimation based on the HSV color space, which is then accompanied by another variational-based estimation of both the reflectance and illumination. Finally, we bring a new perspective to the Retinex model by reconstructing the normal-light image with a novel transformation map given by the estimated reflectance and illumination, which we call radiance mapping. Extensive experiments show that our method obtains outstanding results, both numerically and visually, compared to state-of-the-art methods.