Progressive feature-aware recurrent net for low-light image enhancement.

Mixed mapping across illumination and color domains directly from low-light images is prone to cause domain ambiguity and mislead the holistic enhancement process. The results appear color distorted and unbalanced between overexposure and underexposure. To deal with this issue, we propose a progressive feature-aware recurrent network decoupling the enhancement task into domain-specific subtasks, which gradually restores the illumination and color information of the low-light image under different attention patterns. Specifically, the illumination is adaptively enhanced by the recurrent sub-network with pixel-wise attention. Then, under the guidance of normal illumination, the color enhancement sub-network uses channel attention to balance the restoration of color channels in HSV space. Finally, the preliminary restoration image is further enhanced by the recurrent refined sub-network with dual attention. This progressive learning pattern enables each sub-network to focus on the learning of the specific feature, and the whole network can realize the cumulative learning of multiple features. Experiments on public low-light image datasets demonstrate that our method outperforms the SOTA methods.