ACE-Flow: Auto Color Encoding for Enhanced Low-Light Image Restoration

Jiachen Qiu, Yushen Zuo, Kin-Man Lam

Abstract: Low-light image enhancement is a classic problem in low-level vision tasks, aiming to improve the quality of images captured in poor-lighting scenarios. Conventional deep enhancement models often produce distorted content (e.g., deviated lighting conditions, color bias) in extremely dark regions because they fail to capture comprehensive color information during the reconstruction process. To address these issues, we propose a novel normalizing flow-based model that incorporates an auto-color encoding method, called ACE-Flow, for low-light image enhancement. By leveraging auto-color encoding, our method can encode color information during feature extraction and effectively restore the corrupted image content in challenging regions. Furthermore, our approach can accurately learn the mapping from low-light images to high-quality ground-truth images, because the invertibility property of the normalizing flow implicitly regularizes the learning process. Experiments demonstrate that our method significantly outperforms other promising low-light enhancement models in terms of reconstruction and perceptual metrics. Additionally, the enhanced images produced by our model exhibit rich details with minimal distortion, resulting in superior visual quality.