A Two-Phase Reference-Free Approach for Low-Light Image Enhancement

Reference-free low-light image enhancement methods only employ low-light images during training, thereby significantly alleviating the over-reliance on obtaining paired or unpaired datasets. Existing reference-free low-light image enhancement approaches still struggle to strike a balance between enhancing vivid color and suppressing noise in low-light images. To mitigate such issues, we propose a novel deep learning-based reference-free method that contains two phases, separating the low-light image enhancement into decomposition and refinement problems. In the decomposition phase, we present a value channel prior based on histogram equalization on HSV color space, termed as V-HE prior. Inspired by retinex theory, V-HE prior guides the decomposition network (Dec-Net) to estimate the reflectance component of the value channel. To further refine the pre-enhanced result, we construct a structure-aware loss to guide the refinement network (Ref-Net) in the refinement phase. We conduct extensive experiments to verify the effectiveness of the proposed method, qualitatively and quantitatively. Compared with other reference-free algorithms, our approach effectively addresses the challenges of low-light image enhancement and significantly improves image quality.