Enhanced Illumination-Reflectance Decomposition for Image Enhancement Based on Retinexformer

초록

Enhancing low-light images involves not only improving visibility and contrast but also addressing various degradation artifacts commonly found in poorly lit environments. Many state-of-the-art approaches grounded in Retinex theory have successfully enhanced image quality by recovering reflectance and optimizing illumination. Retinexformer, for instance, tackles the limitations of CNN-based frameworks - such as limited receptive fields and challenges in modeling global context - by employing an illumination-aware Trans-former within a unified architecture. However, existing de-composition techniques often suffer from residual content leakage into the illumination component. To resolve this, we propose a novel decomposition network that explicitly disentangles structure-agnostic illumination and content-preserving reflectance by capturing their mutual interactions. This pro-motes finer restoration in localized areas. Our extensive experiments across diverse benchmark datasets demonstrate that the proposed method consistently outperforms current methods in both quantitative metrics and visual quality. © 2025 IEEE.

키워드