A robust structure and texture aware model for image Retinex

Abstract

Retinex theory is developed mainly to decompose an image into the product of illumination and reflectance based on local derivatives, in which smaller derivatives are given to the illumination and larger derivatives are attributed to the changes in the reflectance. In this paper, we propose a Retinex decomposition model that classifies image derivatives into illumination and reflectance components via structure/texture aware maps. First, we introduce a novel measure called maximum neighbor difference (MND), and utilize it to generate robust structure aware map (RSAM) and robust texture aware map (RTAM) by Huber loss weight from robust statistics. Then, we use RSAM and RTAM as weighting matrices to construct regularization terms of illumination and reflectance, respectively, and obtain a robust structure and texture aware Retinex (RSTAR) model for Retinex decomposition. Finally, we design an alternatively updated algorithm to solve RSTAR, in which the original nonlinear minimization problem is transformed into two linear sub-problems. The performance of RSTAR is evaluated subjectively and objectively on four image datasets in terms of image decomposition/low-light image enhancement, in comparison to eight state-of-the-art Retinex decomposition models. Evaluation results have demonstrated the superior performance of RSTAR in low-light image enhancement.