.png){kind=icon}
.png){kind=icon}
Return
Frequency-conditional diffusion model for low-light image enhancement
.png){kind=icon}
PRE
AI
DOI：10.1016/j.neucom.2025.131383.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Low-light image enhancement (LLIE) aims to improve the visibility and quality of low-light images. The latest developments in diffusion model-based low-light enhancement have shown improvements in generating realistic, detailed images through iterative denoising. However, diffusion models often produce structural incompleteness and blurred contours due to their inability to effectively optimize the frequency–domain characteristics of noise. Specifically, existing methods fail to resolve the inconsistencies between low-frequency image features and noise components, leading to structurally incomplete restorations. Additionally, high-frequency details are easily disturbed by noise during restoration, leading to texture degradation. To address these problems, we propose a frequency-conditional diffusion model for low-light image enhancement. In the low-frequency domain, we employ Two-Dimensional Discrete Wavelet Transform (2D-DWT) to align noise and image features, maintaining the integrity of the image’s structural information. And in order to tackle detail reconstruction challenges, we introduce the Multi-directional Sparse High-frequency Reconstruction Module (MSHR). This module employs a learnable top-k operator to retain key values and suppress low-correlation interference, while integrating horizontal, vertical and diagonal details to enhance high-frequency detail reconstruction. Our method achieves excellent results on mainstream benchmark datasets, such as LOLv1 and LOLv2. Visualization results demonstrate that our method performs better in terms of structural and detail features. The source code is available at https://github.com/JunhaoWang17/Frediff .
Keywords:
low-light image enhancement
diffusion models
frequency-domain optimization
structural completeness
high-frequency detail reconstruction
Journal
IF:
6.5
Papers:
2.5W
Citations:
6.5W
Organization
View more organizations