Towards Color-Faithful Low-Light Image Enhancement via Adaptive Color Debiasing and Saturation Rectification

Low-light image enhancement (LLIE) often results in images that are bright enough but suffer from distorted colors, such as global color shifts and local saturation abnormalities. The authors investigate why current methods fail to restore faithful colors, identifying that the embedded color bias in low-light inputs is not removed by standard color space transformations and instead propagates through the enhancement pipeline.

To solve this, the authors introduce CAGE, a framework that uses a custom color space called AdaLAB. AdaLAB is a cylindrical version of the CIELab space that separates lightness from chrominance, providing a structured basis for color correction. The core of the method is the Adaptive Cylindrical Color Transform (AdaCCT), which performs two key operations:

1. Forward Transform: Before the image is passed to the enhancement backbone, AdaCCT uses learned, image-adaptive parameters to shift hues and scale chroma, effectively suppressing the embedded color bias.
2. Inverse Transform: After the backbone restores the image, AdaCCT reverses the scaling and performs out-of-gamut lightness compensation, which converts color values that would otherwise be clipped into lightness gains, preserving visual detail and preventing saturation artifacts.

This research demonstrates that color restoration in low-light imaging is not merely a byproduct of brightness recovery but requires explicit, adaptive correction of chromatic distributions. By treating color correction as a plug-and-play component that can be integrated into various existing LLIE backbones, the CAGE framework provides a flexible way to improve the visual quality and color accuracy of enhanced images without needing to redesign the core enhancement models themselves.