You Only Need One Color Space: An Efficient Network for Low-light Image Enhancement

TL;DR

The paper tackles color and brightness artifacts in low-light image enhancement by introducing a trainable Horizontal/Vertical-Intensity (HVI) color space and a dual-branch network, CIDNet, that decouples brightness and color. HVI uses a trainable density parameter $k$, hue-perception parameters $\gamma_G$, $\gamma_B$, and a trainable function $T(x)$ to adapt to varying illumination, while CIDNet employs a HV-branch and an intensity-branch with Lighten Cross-Attention and a PHVIT-based reversible mapping to sRGB. The approach achieves state-of-the-art performance across 11 datasets with around 1.88M parameters and 7.57 GFLOPs, demonstrating improved color fidelity, reduced artifacts, and robustness. Overall, this work advances practical LLIE by enabling efficient deployment on edge devices through effective brightness-color decoupling and trainable color space representations.

Abstract

Low-Light Image Enhancement (LLIE) task tends to restore the details and visual information from corrupted low-light images. Most existing methods learn the mapping function between low/normal-light images by Deep Neural Networks (DNNs) on sRGB and HSV color space. Nevertheless, enhancement involves amplifying image signals, and applying these color spaces to low-light images with a low signal-to-noise ratio can introduce sensitivity and instability into the enhancement process. Consequently, this results in the presence of color artifacts and brightness artifacts in the enhanced images. To alleviate this problem, we propose a novel trainable color space, named Horizontal/Vertical-Intensity (HVI). It not only decouples brightness and color from RGB channels to mitigate the instability during enhancement but also adapts to low-light images in different illumination ranges due to the trainable parameters. Further, we design a novel Color and Intensity Decoupling Network (CIDNet) with two branches dedicated to processing the decoupled image brightness and color in the HVI space. Within CIDNet, we introduce the Lightweight Cross-Attention (LCA) module to facilitate interaction between image structure and content information in both branches, while also suppressing noise in low-light images. Finally, we conducted 22 quantitative and qualitative experiments to show that the proposed CIDNet outperforms the state-of-the-art methods on 11 datasets. The code is available at https://github.com/Fediory/HVI-CIDNet.

Figures (12)

• Figure 1: The sensitivity comparison of different color spaces in low-light enhancement. The notations $\Delta$R, $\Delta$V, and $\Delta$I represent a tiny variation in the axis of Red (sRGB), Value (HSV), and Intensity (HVI), respectively. After enhancement processing, noticeable color artifacts can be observed in the sRGB and HSV space results.
• Figure 2: HSV color space visualization. The two black circles in the left image indicate the discontinuous color positions along the hue axis, while the red box in the right image displays a pure black plane with Value$=0$.
• Figure 3: The overview of the proposed CIDNet. (a) HVI Color Transformation (HVIT) takes an sRGB image as input and generates HV color map and intensity map as outputs. (b) Enhancement Network performs the main processing, utilizing a dual-branch UNet architecture, which contains six Lighten Cross-Attention (LCA) blocks. Lastly, we apply Perceptual-inverse HVI Transform (PHVIT) to take a light-up HVI map as input and transform it into an sRGB-enhanced image.
• Figure 4: The dual-branch Lighten Cross-Attention (LCA) block (i.e., I-branch and HV-branch). The LCA incorporates a Cross Attention Block (CAB), an Intensity Enhance Layer (IEL), and a Color Denoise Layer (CDL). The feature embedding convolution layers contains a $1\times1$ depth-wise convolution and a $3 \times 3$ group convolution.
• Figure 5: Visual comparisons of the enhanced results by different methods on LOLv1 and LOLv2. (Zoom in for the best view.)