Source-linked AI summary
Zero-Reference Deep Curve Estimation for Low-Light Image Enhancement
Chunle Guo, Chongyi Li, Jichang Guo, Chen Change Loy, Junhui Hou, Sam Kwong, Runmin Cong
TL;DR
Low-light enhancement traditionally depends on reference images or assumptions that may not fit diverse lighting conditions. Zero-DCE estimates pixel-wise, high-order enhancement curves with a lightweight network and trains them using differentiable non-reference losses. Experiments report superiority over existing methods, broad lighting-condition generalization, and efficient deployment, while the method assumes locally similar pixel intensities and leaves noise and hard semantic cases for future work.
Problem
Low-light enhancement needs to improve poorly illuminated images despite limited suitability of paired or unpaired reference data and varied lighting conditions.
Method
Zero-DCE formulates enhancement as image-specific curve estimation, using DCE-Net to predict pixel-wise high-order curves trained with differentiable non-reference losses.
Results
Zero-DCE is reported to outperform existing methods qualitatively and quantitatively, generalize across lighting conditions, and process 640×480×3 images at about 500 FPS on GPU.
Takeaways & Limitations
The paper demonstrates that low-light enhancement can be trained end-to-end without reference images using task-specific non-reference losses and simple curve mapping.
Takeaways & Limitations
The pixel-wise curve formulation assumes pixels in a local region have the same intensity and adjustment curves; future work will address hard cases and noise.
Abstract
from arXiv · showhide
The paper presents a novel method, Zero-Reference Deep Curve Estimation (Zero-DCE), which formulates light enhancement as a task of image-specific curve estimation with a deep network. Our method trains a lightweight deep network, DCE-Net, to estimate pixel-wise and high-order curves for dynamic range adjustment of a given image. The curve estimation is specially designed, considering pixel value range, monotonicity, and differentiability. Zero-DCE is appealing in its relaxed assumption on reference images, i.e., it does not require any paired or unpaired data during training. This is achieved through a set of carefully formulated non-reference loss functions, which implicitly measure the enhancement quality and drive the learning of the network. Our method is efficient as image enhancement can be achieved by an intuitive and simple nonlinear curve mapping. Despite its simplicity, we show that it generalizes well to diverse lighting conditions. Extensive experiments on various benchmarks demonstrate the advantages of our method over state-of-the-art methods qualitatively and quantitatively. Furthermore, the potential benefits of our Zero-DCE to face detection in the dark are discussed. Code and model will be available at https://github.com/Li-Chongyi/Zero-DCE.
1. Introduction
Zero-DCE reframes low-light enhancement as image-specific curve estimation rather than image-to-image mapping, enabling zero-reference training across diverse lighting conditions. The method uses non-reference losses and a lightweight network to produce efficient enhancement with reported qualitative, quantitative, and real-time advantages.
- Motivation: Low-light photos compromise aesthetic quality and can lead to inaccurate object or face recognition.The motivating conditions include inadequate or unbalanced lighting, extreme back light, and under-exposure.
- Method: Zero-DCE estimates high-order curves from each low-light image and applies them pixel-wise to adjust its dynamic range.The formulation maintains the enhanced image range and preserves neighboring-pixel contrast.
- Training: Zero-reference training avoids paired and unpaired data by using spatial consistency, exposure control, color constancy, and illumination smoothness losses.These non-reference losses account for multiple factors of light enhancement and implicitly evaluate output quality.
- Generalization: The method is reported to generalize across lighting conditions while avoiding overfitting associated with dependence on paired or unpaired training data.The contribution statement links data independence with broader lighting-condition generalization.
- Results: Zero-DCE reportedly surpasses state-of-the-art methods qualitatively and quantitatively, improves face detection potential, and processes 640×480×3 images at about 500 FPS on GPU.Training takes about 30 minutes, according to the introduction.
2. Related Work
Prior low-light enhancement methods rely on histogram manipulation, physical illumination models, or paired and unpaired training data. Zero-DCE instead uses image-specific curve mapping with non-reference losses, eliminating reference-data requirements while targeting efficient, artifact-free enhancement.
- Conventional Methods: Conventional methods enhance illumination through global or local histogram adjustment, or by decomposing images into reflectance and illumination under Retinex theory.Retinex-based approaches commonly assume reflectance remains consistent across lighting conditions.
- Zero-DCE: Zero-DCE replaces histogram or physical-model-based enhancement with image-specific curve mapping designed to avoid unrealistic artifacts.Its related-work comparison presents curve mapping as an alternative to fortuitous histogram changes and potentially inaccurate physical models.
- Data-Driven Methods: Data-driven methods include CNN approaches that usually require paired data and GAN approaches that can eliminate pairing but require carefully selected unpaired data.Paired-data collection can involve light degradation, camera-setting changes, or image retouching.
- Data-Driven Methods: Paired-data methods are costly to scale and may include factitious or unrealistic training data, with constraints reflected in poor generalization to real-world lighting variation.The supplied related-work passage states that these constraints affect performance on real-world images of various light intensities.
- Zero-DCE: Zero-DCE removes paired and unpaired reference-data requirements, uses non-reference losses, and emphasizes efficiency through a lightweight network and curve mapping.The paper identifies these as three advantages over existing data-driven methods.
3. Methodology
Zero-DCE estimates image-specific, pixel-wise high-order curves and applies them iteratively to adjust image dynamic range. DCE-Net learns these curves using non-reference losses that enforce spatial, exposure, color, and illumination properties.
- DCE-Net: DCE-Net estimates pixel-wise curve parameter maps from a low-light image and applies the resulting LE-curves iteratively to produce an enhanced image.The network outputs parameter maps for eight iterations, with three maps per iteration for the RGB channels.
- Light-Enhancement Curve (LE-curve): The quadratic LE-curve constrains enhanced pixel values to [0,1], preserves monotonicity, and remains differentiable for network training.Its parameter α lies in [−1,1] and controls both curve magnitude and exposure level.
- Light-Enhancement Curve (LE-curve): Applying separate curves to RGB channels preserves inherent color and reduces the risk of over-saturation.The estimated channel maps can differ in value while showing similar adjustment tendencies, revealing dark-region content while preserving bright regions.
- Light-Enhancement Curve (LE-curve): Iterative application produces higher-order curves with greater curvature and wider adjustment capability than a single global curve.The paper sets the iteration number n to 8; pixel-wise parameter maps then allow local rather than purely global adjustment.
- Non-Reference Loss Functions: Zero-DCE trains DCE-Net with spatial consistency, exposure control, color constancy, illumination smoothness, and a weighted total loss.These losses preserve neighboring-region differences, control local exposure, correct color deviations, and preserve monotonicity relations through smooth parameter maps.
4. Experiments
Zero-DCE is evaluated through implementation details, ablations, visual and quantitative comparisons, and a face-detection task in dark images. The experiments examine efficiency, parameter and training-data effects, enhancement quality, and downstream detection performance.
- Implementation Details: The model is trained on multi-exposure data, using 2,422 images for training and the remainder for validation after splitting 3,022 Part1 images.Training images are resized to 512×512; the framework uses PyTorch and an NVIDIA 2080Ti GPU.
- Ablation Study: Zero-DCE7−32−8 provides the selected trade-off between efficiency and restoration performance, while one iteration has limited adjustment capability.Three convolutional layers already produce satisfactory results, whereas seven layers with 32 feature maps and 8 or 16 iterations are visually strongest.
- Ablation Study: Removing over-exposed training data causes Zero-DCE to over-enhance well-lit regions, demonstrating that training-data exposure diversity affects enhancement behavior.The ablation compares low-light-only, larger low-light, and multi-exposure training sets.
- Visual and Perceptual Comparisons: Zero-DCE produces natural exposure and clear details in challenging back-lit faces, while competing methods exhibit unclear faces, over-exposure, oversmoothing, noise, or color deviation.In an indoor example, the method enhances dark regions while preserving input colors without obvious noise or color casts.
- Quantitative Comparisons: Zero-DCE achieves the highest average User Study score across 202 testing images and the best values on all reported full-reference metrics in Table 2.The evaluation includes NPE, LIME, MEF, DICM, and VV image sets, with PSNR, SSIM, and MAE used for full-reference assessment.
- Face Detection in the Dark: After enhancement, DSFD precision increases considerably on dark images, with Zero-DCE performing better than RetinexNet in the high-recall area.The face-detection evaluation uses 6,000 training and validation images from DARK FACE.
5. Conclusion
Zero-DCE performs low-light enhancement with a deep network trained without reference images by estimating image-specific curves and using differentiable non-reference losses. Experiments show superiority over existing enhancement methods, while future work targets hard cases and noise.
- Zero-DCE is trained end-to-end without reference images by formulating enhancement as image-specific curve estimation.
- Differentiable non-reference losses enable the network to learn the enhancement process without paired or unpaired training data.
- Experiments demonstrate that Zero-DCE outperforms existing light enhancement methods.
- Future work will address hard cases using semantic information and examine the effects of noise.