.png){kind=icon}
.png){kind=icon}
Return
Multiscale hybrid feature guided normalizing flow for low-light image enhancement
.png){kind=icon}
PRE
AI
DOI：10.1016/j.compeleceng.2024.109922.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Conditional normalizing flow (CNF) can perform reversible transformation to learn the distribution of the normal-light image guided by conditional feature from the low-light image. This innovative generative model presents a unique solution for low-light image enhancement. However, existing CNF-based models completely adopt CNN architectures to extract the conditional feature, which only concentrates on the representation of local information. Besides, conditional affine coupling (CAC) layer in CNF merely executes reversible transformation on the part of the feature, affecting the transformation capability of the overall model. The above limitations have a significant impact on the performance of CNF. Therefore, this paper proposes a novel and powerful CNF-based model named multiscale hybrid feature guided normalizing flow (MHFlow) to stimulate the potential of CNF in low-light image enhancement. Specifically, MHFlow consists of a conditional encoder and an invertible network. In the conditional encoder, we design multiscale hybrid feature encode blocks (MHEB) to extract the multiscale local and global information and conduct the feature fusion based on the cross-attention mechanism. In the invertible network, we construct overlapped conditional affine coupling (OCAC) to perform sufficient transformation for the flow feature, which enhances the transformation capability of the invertible network. Extensive experiments demonstrate that our proposed MHFlow shows better performance than current state-of-the-art (SOTA) methods based on the per-pixel reconstruction in terms of quantitative evaluation and visual quality. The source code and pre-trained models are available at https://github.com/huyin-NJUPT-IPR/MHFlow.
Keywords:
Low-light image enhancement
Multiscale hybrid feature
Conditional normalizing flow
Overlapped conditional affine coupling
Journal
C
IF:
4.9
Papers:
6.7K
Citations:
1.3W