Source-linked AI summary
Unsupervised Night Image Enhancement: When Layer Decomposition Meets Light-Effects Suppression
Yeying Jin, Wenhan Yang, Robby T. Tan
TL;DR
Night enhancement must address both dark regions and uneven bright light effects, while paired real training data and physically correct synthetic rendering are difficult to obtain. This paper introduces an unsupervised unified framework combining layer decomposition with guided light-effects suppression, and reports that it outperforms state-of-the-art methods on real images.
Problem
Night-image enhancement methods mainly brighten dark regions, but light effects can be amplified; paired data and physically correct synthetic rendering are difficult to obtain.
Method
An unsupervised framework decomposes shading, reflectance, and light-effects layers, then uses guided unpaired suppression and structure/high-frequency consistency losses.
Results
The method outperforms state-of-the-art visibility-enhancement and light-effects-suppression methods on real images.
Takeaways & Limitations
The approach jointly suppresses night light effects and boosts dark-region intensity from a single night image.
Takeaways & Limitations
Because the layer model assumes a linear gamma function, training with nonlinear images yields background scenes that approximate physically correct values.
Abstract
from arXiv · showhide
Night images suffer not only from low light, but also from uneven distributions of light. Most existing night visibility enhancement methods focus mainly on enhancing low-light regions. This inevitably leads to over enhancement and saturation in bright regions, such as those regions affected by light effects (glare, floodlight, etc). To address this problem, we need to suppress the light effects in bright regions while, at the same time, boosting the intensity of dark regions. With this idea in mind, we introduce an unsupervised method that integrates a layer decomposition network and a light-effects suppression network. Given a single night image as input, our decomposition network learns to decompose shading, reflectance and light-effects layers, guided by unsupervised layer-specific prior losses. Our light-effects suppression network further suppresses the light effects and, at the same time, enhances the illumination in dark regions. This light-effects suppression network exploits the estimated light-effects layer as the guidance to focus on the light-effects regions. To recover the background details and reduce hallucination/artefacts, we propose structure and high-frequency consistency losses. Our quantitative and qualitative evaluations on real images show that our method outperforms state-of-the-art methods in suppressing night light effects and boosting the intensity of dark regions.
1 Introduction
Night images combine dark regions with uneven bright light effects, which ordinary low-light enhancement can intensify. The paper proposes an unsupervised unified framework that decomposes layers, suppresses light effects, enhances dark regions, and preserves background details.
- Night images contain uneven illumination, with dark areas alongside significantly brighter regions caused by light effects.
- Low-light enhancement methods can amplify night light effects and further impair visibility instead of handling both brightness and light effects.
- Paired night images with and without light effects are difficult to collect, while physically correct synthetic light-effects rendering remains challenging.
- The proposed unsupervised framework integrates layer decomposition with light-effects suppression, decomposing inputs into shading, reflectance, and light-effects layers.
- Structure and high-frequency consistency losses preserve background details and reduce hallucination and unwanted artefacts.
- Experiments show quantitative and qualitative superiority over state-of-the-art methods for suppressing light-effects regions and enhancing dark regions.
2 Related Work
Prior work includes light-effects suppression, night dehazing, and low-light enhancement, but these approaches address different subsets of the problem. The proposed method targets simultaneous light-effects suppression and dark-region enhancement.
- Sharma and Tan use camera response function estimation and HDR imaging to suppress light effects and improve night-image dynamic range.
- Night dehazing methods suppress glow in foggy or hazy nights, but they are not designed for light-effects removal in clear night images.
- Low-light enhancement methods boost brightness without considering night light effects, and many deep-learning methods require paired low- and normal-light images.
- The proposed framework differs by integrating decomposition and light-effects suppression to enhance dark regions while suppressing light effects simultaneously.
3 Proposed Method
The proposed framework decomposes a night image into shading, reflectance, and light-effects layers, then uses those layers to guide unsupervised light-effects suppression. Layer-specific priors, attention guidance, adaptive fusion, and structure/HF consistency losses support decomposition, region-focused refinement, and background-detail preservation.
- Model-Based Layer Decomposition Network: The image-layer model represents the input as I = R ⊙ L + G, with R, L, and G denoting reflectance, shading, and light-effects layers.The model assumes a linear gamma function and guides network design rather than being used explicitly during inference.
- Model-Based Layer Decomposition Network: Three separate networks estimate the light-effects, shading, and reflectance layers from the input using unsupervised losses.The networks are Light-Effects-Net φG, Shading-Net φL, and Reflectance-Net φR.
- Model-Based Layer Decomposition Network: Initialization uses a channel-maximum shading map and a Laplacian-derived light-effects map to reduce decomposition ambiguity.The light-effects initialization exploits the smooth-variation characteristic of light effects.
- Model-Based Layer Decomposition Network: Gradient exclusion separates the smooth light-effects layer from the background scene in gradient space, while reconstruction and color-constancy losses constrain the decomposition.The background scene is Jinit = R ⊙ L; reconstruction recombines estimated layers to recover the input, and color constancy balances channel intensities.
- Light-Effects Suppression Network: The suppression network refines Jinit into Jrefine using G as guidance, modulating multiscale feature maps so processing focuses on light-effects regions.The generator receives the concatenation of Jinit and G, while an auxiliary classifier distinguishes light-effects and unpaired light-effects-free domains.
- Light-Effects Suppression Network: Adaptive RGB fusion produces Igray for VGG structure features and guided-filter HF features that are less affected by light effects and help preserve background details.The channel weights are low for under-exposed or light-effects pixels, and the consistency losses are also intended to reduce hallucination.
4 Experimental Results
Experiments show that the integrated framework suppresses night light effects while enhancing dark regions, with joint training and its key components improving qualitative results. Quantitative evaluations report strong performance on night and low-light datasets.
- Evaluation: 210 outputs were rated by 12 participants across realism, light-effects suppression, and visibility criteria in the user study.The study sampled 30 outputs per method across seven methods.
- Quantitative results: Our method achieved the highest PSNR and SSIM scores on the night-data quantitative evaluation.Table 2 reports the comparison among light-effects suppression methods.
- Baseline comparisons: Low-light enhancement baselines wrongly intensify light effects, whereas night dehazing baselines are too dark because they are not designed to enhance dark regions.These comparisons concern real night-image enhancement with both light effects and dark regions.
- Low-light enhancement: Our method achieved better PSNR and SSIM performance than baseline methods on LOL-test and LOL-Real, with results showing generality across low-light datasets.The LOL dataset contains 485 training and 15 testing images, while LOL-Real includes 100 testing images with more diversified scenes.
- Joint training: Joint suppression and dark-region boosting outperform suppression alone, boosting alone, and sequentially trained alternatives.Suppressing first without joint training enhances artefacts and residual light effects, while boosting first causes information and detail loss from amplified light effects.
- Ablation studies: Ablations show that layer decomposition, light-effects guidance, and structure and HF-features consistency losses each improve the framework’s results.Guidance helps distinguish and suppress light-effects regions, while consistency losses suppress artefacts and restore details.
5 Conclusion
The paper proposes an integrated network that suppresses light effects and boosts dark regions from a single night image. Experiments report that it outperforms state-of-the-art visibility-enhancement and light-effects-suppression methods.
- Conclusion: The proposed method jointly suppresses light effects and boosts dark-region intensity from a single night image.It casts light-effects suppression as an unsupervised decomposition problem and integrates layer decomposition with light-effects suppression.
- Conclusion: Experiments show that the method outperforms state-of-the-art visibility-enhancement and light-effects-suppression methods.The conclusion summarizes the reported experimental comparison.