Fractional structure and texture aware model for image Retinex and low-light enhancement

This paper proposes a fractional structure and texture aware Retinex (FSTAR) model for image decomposition with application to low-light enhancement. First, a novel structure aware measure called Maximum Fractional Difference (MFD) is introduced, which is the maximum of fractional differences of the input image in eight symmetric directions. Then fractional structure and texture aware maps are built based on MFD and Huber function from robust statistics, which are used as weighted matrices in the regularization terms of reflectance and illumination components. To ensure the intrinsic physical constraints of Retinex on two components, two extra penalty terms are incorporated into the objective function of FSTAR to penalize deviations of the estimated components from the corresponding constraints. The FSTAR is solved via an alternating iterative algorithm; i.e., the objective function is minimized with respect to one variable with another variable fixed, and this process is done alternately until termination condition is met. Qualitative and quantitative evaluations of FSTAR on three low-light image datasets, compared to ten state-of-the-art methods, show its superior performance in Retinex decomposition and low-light image enhancement.