CSDM: A Cross-Scale Decomposition Method for Low-Light Image Enhancement

Abstract

Inadequate lighting condition is one of the main factors affecting image quality. For high-quality image display and vision applications, low-light image enhancement is an extremely challenging task. However, most existing methods are difficult to remove heavy noise in the low-light enhancement process effectively. To solve this problem, this paper presents a novel cross-scale decomposition method (CSDM) for low-light image enhancement. The main idea is to capture the small-scale texture and heavy noise, and preserve the large-scale structure by designing a Gaussian regularization based on a cross-scale relative relationship. An optimization objective function is established to transform image decomposition, low-light enhancement and noise suppression into a unified optimization framework. The framework uses sequential decomposition to estimate texture and structure layers. Firstly, the structure layer of the image is estimated, and an illumination prior is introduced to improve the contrast. Then the texture layer is estimated, and a low-rank prior is applied to eliminate noise. Finally, the structure and texture layers are combined to obtain a final enhanced image. Experiment results on public datasets and an industrial site demonstrate that the advantages of our method from both qualitative and quantitative perspectives.