TY - JOUR
T1 - DCWNet
T2 - Bridging Dual-Codebook Reconstruction and Wavelet Refinement for Low-Light Image Enhancement
AU - Shi, Kangbiao
AU - He, Zhaokun
AU - Doan, Anh Dzung
AU - Yan, Li
AU - Yan, Qingsen
N1 - Publisher Copyright:
© 1991-2012 IEEE.
PY - 2026
Y1 - 2026
N2 - Low-Light Image Enhancement (LLIE) aims to improve the visual quality of images degraded by poor illumination. Existing LLIE methods often apply uniform enhancement across the entire image, ignoring the intricate coupling between illumination and color, which leads to color distortion and uneven exposure. Recent HVI-based approaches decouple illumination and color, but they still struggle in extremely dark regions, where the chroma signal is severely attenuated and dominated by noise, resulting in reduced color discriminability and residual artifacts. To address these challenges, we introduce DCWNet, a novel two-branch LLIE framework in the HVI color space. Specifically, the HV-branch reconstructs the HV color map using a dual-codebook strategy: a Bright Codebook for stable, natural-light colors, and a Dark Codebook that adaptively captures chromatic distributions in dark regions. The Context-Aware Module further enhances spatial consistency, while an uncertainty-guided fusion mechanism focuses on high-entropy regions, jointly suppressing noise and correcting global color bias. In the I-branch, a Frequency-Decoupled Wavelet Transform module decomposes the intensity map into high- and low-frequency components for separate processing. By preserving fine details and texture information, it effectively alleviates local overexposure and underexposure through region-aware brightness optimization. Extensive experiments on ten benchmark datasets demonstrate that DCWNet outperforms state-of-the-art LLIE methods.
AB - Low-Light Image Enhancement (LLIE) aims to improve the visual quality of images degraded by poor illumination. Existing LLIE methods often apply uniform enhancement across the entire image, ignoring the intricate coupling between illumination and color, which leads to color distortion and uneven exposure. Recent HVI-based approaches decouple illumination and color, but they still struggle in extremely dark regions, where the chroma signal is severely attenuated and dominated by noise, resulting in reduced color discriminability and residual artifacts. To address these challenges, we introduce DCWNet, a novel two-branch LLIE framework in the HVI color space. Specifically, the HV-branch reconstructs the HV color map using a dual-codebook strategy: a Bright Codebook for stable, natural-light colors, and a Dark Codebook that adaptively captures chromatic distributions in dark regions. The Context-Aware Module further enhances spatial consistency, while an uncertainty-guided fusion mechanism focuses on high-entropy regions, jointly suppressing noise and correcting global color bias. In the I-branch, a Frequency-Decoupled Wavelet Transform module decomposes the intensity map into high- and low-frequency components for separate processing. By preserving fine details and texture information, it effectively alleviates local overexposure and underexposure through region-aware brightness optimization. Extensive experiments on ten benchmark datasets demonstrate that DCWNet outperforms state-of-the-art LLIE methods.
KW - Bright Codebook
KW - Dark Codebook
KW - HVI color space
KW - low-light image enhancement
KW - wavelet transform
UR - https://www.scopus.com/pages/publications/105048607723
U2 - 10.1109/TCSVT.2026.3723107
DO - 10.1109/TCSVT.2026.3723107
M3 - 文章
AN - SCOPUS:105048607723
SN - 1051-8215
JO - IEEE Transactions on Circuits and Systems for Video Technology
JF - IEEE Transactions on Circuits and Systems for Video Technology
ER -