Low-Light Image Enhancement via Regularized Gaussian Fields Model

Retinex decomposition is a prevalent solution to low-light image enhancement. It is usually considered as a constrained optimization problem. To improve enhancement performance, the Retinex model is incorporated with various prior constraints, which make the optimization process complicated and difficult. In this paper, a method of low-light enhancement with regularized Gaussian Fields (RGF) model is proposed to address this issue. Firstly, we construct an RGF-based optimization model to formulate simultaneous reflectance and illumination estimation as an unconstrained optimization problem. Therefore, Retinex decomposition in RGF model can be solved by gradient descent techniques and has superiority on computational convenience. Then, to suppress noise and preserve detail in the estimated reflectance, the detail-preserving model based on Gaussian total variation (GTV) is established. The qualitative and quantitative comparisons on several public datasets demonstrate the superiority of our method over several state-of-the-arts in terms of enhancement efficiency and quality.