A Variational Retinex Model With Structure-Awareness Regularization for Single-Image Low-Light Enhancement

Low-light image enhancement (LLIE) is a method of improving the visual quality of images captured in weak illumination conditions. In such conditions, the images tend to be noisy, hazy, and have low contrast, making them difficult to distinguish details. LLIE techniques have many practical applications in various fields, including surveillance, astronomy, medical imaging, and consumer photography. The total variational method is a sound solution in this field. However, requirement of an overall spatial smoothness of the illumination map leads to the failure of recovering intricate details. This paper proposes that the interaction between the global spatial smoothness and the detail recovery in the total variational Retinex model can be optimized by adopting a structure-awareness regularization term. The resultant non-linear model is more effective than the original one for LLIE. As a model-based method, its performance does not rely on architecture engineering, super-parameter tuning, or specific training dataset. Experiments of the proposed formulation on various challenging low-light images yield promising results. It is shown that this method not only produces visually pleasing pictures, but it is also quantitatively superior in that the calculated full-reference, no-reference, and semantic metrics are beyond most of state-of-the-art methods. It has a better generalization capability and stability than learning-based methods. Due to its flexibility and effectiveness, the proposed method can be deployed as a pre-processing subroutine for high-level computer vision applications.