The article presents a novel approach to Low-Light Image Enhancement (LLIE) using a new trainable color space, HVI, and a decoupling network, CIDNet. Existing methods using sRGB and HSV color spaces often introduce instability and sensitivity, leading to color and brightness artifacts in enhanced images. The proposed HVI color space decouples brightness and color from RGB channels, making it more adaptable to low-light images. The CIDNet, with its Lightweight Cross-Attention module, further aids in processing the decoupled image brightness and color in the HVI space, reducing noise and improving overall image quality. The proposed approach showed superior performance over existing methods in multiple experiments.
Publication date: 9 Feb 2024
Project Page: https://github.com/Fediory/HVI-CIDNet
Paper: https://arxiv.org/pdf/2402.05809
