Robust Structure-Aware Low-Light Enhancement via a Latent Low-Rank Embedding Retinex Model

Robustness to noise is a significant issue in the field of low-light image enhancement (LLIE). Most existing LLIE methods, which ignore the intensive noise contained in images or use a pre/postprocessing step to remove noise, fail to achieve pleasant visual effects in practical tasks. To address this problem, in this paper, based on Retinex theory, we propose jointly integrating low-light enhancement and denoising into a unified framework with latent low-rank embedding. Specifically, an ℓ
2,1
-norm term is first introduced to measure the reconstruction error and enhance the robustness of decomposed components to various types of noise. Then, a smooth prior is applied to the illumination component, and a structure-aware gradient fidelity term and latent low-rank embedding are deployed to resist noise and retain the texture of the reflectance component. Considering the expensive time cost of low-rank processing, we design an accelerated low-rank approximation strategy without losing structural information. Afterward, the decomposed components are sequentially estimated, avoiding noise from the other component. Finally, the enhanced images are generated by combining the estimated reflectance with the illumination adjusted based on gamma correction. Extensive experiments demonstrate that our method performs significantly better than several state-of-the-art methods in terms of enhancing and denoising on three public datasets.