Abstract
Low-light image enhancement (LIE) remains a significant challenge due to the complex interplay between contrast restoration and noise suppression. Traditional supervised methods are limited by the difficulty of acquiring high-quality ground truth pairs, while unsupervised methods often rely on rigid, handcrafted priors that fail to generalize. This paper introduces PairLIE, an unsupervised framework that leverages pairs of low-light images of the same scene to learn adaptive priors. By enforcing reflectance consistency across varying low-light levels, PairLIE achieves state-of-the-art performance with a remarkably simple architecture, effectively bridging the gap between supervised and unsupervised learning.
Introduction
Enhancing images captured in low-light conditions is essential for various downstream tasks, yet it is inherently difficult because sensor noise often overwhelms the underlying signal. Current research typically follows one of two paths: supervised learning, which requires perfectly aligned low-light and normal-light pairs, or unsupervised learning, which utilizes mathematical assumptions (priors) about image properties like illumination smoothness.
As illustrated in Figure 1, single-image methods (a) rely heavily on fixed priors, which limits their adaptability. In contrast, the proposed PairLIE framework (b) utilizes paired low-light instances—which are significantly easier to capture than normal-light references—to learn adaptive priors directly from the data. This approach allows the model to better handle the diversity of real-world lighting and textures.
Methodology
PairLIE is grounded in Retinex Theory, which posits that an image ($I$) can be decomposed into an illumination map ($L$) and a reflectance map ($R$): $I = L \circ R$. The core hypothesis is that while illumination varies between two shots of the same scene, the physical reflectance remains constant.
Network Architecture
The framework consists of three primary components:
- P-Net (Projection): A preprocessing network that projects the raw input into a cleaner feature space, effectively separating noise and non-Retinex artifacts from the core signal.
- L-Net & R-Net: Two simple five-layer convolutional neural networks (CNNs) dedicated to estimating the illumination and reflectance components, respectively.
Loss Functions and Consistency
The training process is driven by the Reflectance Consistency Loss ($\mathcal{L}_C$). Given two low-light images ($I_1, I_2$) of the same scene, the network is penalized if the predicted reflectance maps ($R_1, R_2$) differ. This constraint forces the model to ignore transient noise (which differs between shots) and focus on the invariant scene properties. Additional losses include a Projection Loss ($\mathcal{L}_P$) to guide the P-Net and a Reconstruction Loss ($\mathcal{L}_R$) to ensure the decomposed components can accurately recreate the original input.
During inference, the model requires only a single image. The illumination map is adjusted via a scalar factor ($\lambda$) before being recombined with the reflectance map to produce the final enhanced output.
Experiments and Results
The authors evaluated PairLIE on the LOL (Low-Light) and SICE datasets. For training, only the low-light portions of these datasets were used, maintaining the unsupervised nature of the approach.
Quantitative Performance
PairLIE demonstrates a significant performance leap over existing unsupervised and zero-shot methods. On the LOL benchmark, PairLIE achieved a PSNR of 19.51, substantially outperforming the widely used Zero-DCE (14.86). On the SICE dataset, it reached a PSNR of 21.32, proving its robustness across different multi-exposure scenarios.
Qualitative Analysis
Visual results indicate that PairLIE is exceptionally effective at suppressing color noise in deep shadows. While competing methods often introduce blotchy artifacts or unnatural color shifts when boosting dark regions, PairLIE maintains cleaner textures and more natural color reproduction. This is attributed to the reflectance consistency constraint acting as a natural, self-supervised denoiser.
Discussion and Conclusion
PairLIE proves that the strength of a learning constraint can be more impactful than the complexity of the network architecture. By using simple five-layer CNNs and a clever paired-data strategy, the model achieves high efficiency and competitive accuracy.
However, the method does have limitations. It assumes that training pairs are well-aligned; significant camera movement or dynamic objects between shots can degrade the consistency loss. Future work may explore synthetic pair generation from single images to remove the requirement for physical pairs during training, as well as deeper exploitation of the relationship between illumination maps in a sequence. Despite these challenges, PairLIE represents a significant step forward in making high-quality low-light enhancement more accessible and computationally efficient.
Summary updated