Temel Kavramlar
A novel trainable color space called Horizontal/Vertical-Intensity (HVI) is proposed to decouple image brightness and color, enabling efficient low-light image enhancement.
Özet
The paper introduces a novel color space called Horizontal/Vertical-Intensity (HVI) that decouples image brightness and color information. The HVI color space has trainable parameters that allow it to adapt to different low-light conditions.
Based on the HVI color space, the authors propose a dual-branch network called Color and Intensity Decoupling Network (CIDNet) that separately processes the brightness and color information. CIDNet uses a Lightweight Cross-Attention (LCA) module to facilitate interaction between the brightness and color branches, allowing them to complement each other.
The authors conduct extensive experiments on 11 datasets and show that CIDNet outperforms state-of-the-art low-light image enhancement methods in terms of both quantitative and qualitative metrics. CIDNet is also shown to be efficient in terms of model size and computational complexity.
The key highlights of the paper are:
- Introduction of the HVI color space that decouples brightness and color information and can adapt to different low-light conditions.
- Proposal of the CIDNet architecture that leverages the HVI color space for efficient low-light image enhancement.
- Design of the LCA module to enable effective interaction between the brightness and color branches of CIDNet.
- Comprehensive evaluation on multiple datasets demonstrating the superior performance and efficiency of CIDNet compared to state-of-the-art methods.
İstatistikler
The paper reports the following key metrics:
PSNR (Peak Signal-to-Noise Ratio)
SSIM (Structural Similarity Index)
LPIPS (Learned Perceptual Image Patch Similarity)
BRISQUE (Blind/Referenceless Image Spatial Quality Evaluator)
NIQE (Naturalness Image Quality Evaluator)
FLOPs (Floating-point Operations)
Number of Parameters
Alıntılar
"To sort out the aforementioned color space challenges, we first present a trainable Horizontal/Vertical-Intensity (HVI) color space."
"Based on the HVI color space, we propose a novel dual-branch network, CIDNet, to concurrently process the brightness and color of low-light images."
"We design a bidirectional LCA to facilitate interaction between the HV-branch and Intensity-branch, allowing the scene information in each branch to complement and improve the visual effects of the enhanced image."