Low-Light Image Enhancement via Adaptive Shape and Texture Prior

Low light images affect various computer vision algorithms due to their low visibility and much noise hidden in dark regions. Although many methods based on the Retinex theory, which decomposes an observed image into the reflectance and illumination, have been proposed to alleviate the problem, existing methods inevitably cause under-and over-enhancement. In this paper, we propose a new joint optimization equation that sufficiently considers the features of both reflectance and illumination. More concretely, we adopt L2-Lp norm regularization terms to estimate the reflectance as much as possible to preserve details and textures, and the illumination as much as possible to preserve the structure information with texture-less. We solve the optimization equation in an alternating minimization method. Furthermore, we introduce a new adaptive texture prior to reveal more details and textures with noise reduction on both bright and dark regions. Experimental results, including qualitative and quantitative evaluations, show that the proposed method can establish a better performance than the other state-of-the-art methods.