.png){kind=icon}
.png){kind=icon}
Return
SEEGAN: signal-to-noise ratio-guided low-light image enhancement with GAN
.png){kind=icon}
PRE
AI
A
M
T
J
B
DOI：10.1117/1.JEI.35.2.023012.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
In recent years, numerous excellent low-light image enhancement (LLIE) algorithms have been proposed. However, most struggle with exposure control under complex lighting conditions, typically manifesting as over-exposure in bright regions and under-exposure or noise amplification in dark regions. We propose the signal-to-noise ratio (SNR)-guided dual-generator framework (SDF), a generative adversarial network-based framework for unsupervised LLIE that balances signal enhancement, noise suppression, and feature preservation. SDF integrates two generators fed with inputs of different SNRs and optimizes a global discriminator alongside a visual geometry group-based local feature discriminator: the SNR difference between inputs drives mutual constraint of generators to adapt to real data distribution, whereas the dual discriminators jointly ensure enhanced images retain both dark-region and bright-region features. As the core generator of SDF, signal-to-noise ratio-guided efficient enhancement generative adversarial network consists of two key modules-long-short range feature extraction and long-short range feature reconstruction-both built on our proposed long-short range convolutions. By leveraging luminance-aware information and nonlocal features extracted via long-range convolutions to guide short-range convolutions, LSConv achieves precise feature reconstruction for uniform exposure adjustment. Comprehensive evaluations on benchmark datasets show that our method outperforms state-of-the-art approaches in LLIE tasks and demonstrates strong robustness in downstream low-light vision applications.
Keywords:
low-light enhancement
unsupervised learning
generative adversarial networks
balanced exposure
dual discriminators
Journal
J
IF:
1
Papers:
109
Citations:
2.7K
Organization
Cited Papers