[论文解读] Zero-Reference Deep Curve Estimation for Low-Light Image Enhancement

Chunle Guo, Chongyi Li+5 more|arXiv (Cornell University)|Jan 19, 2020

Image Enhancement Techniques参考文献 36被引用 141

一句话总结

Zero-DCE 在不使用任何参考图像的前提下进行训练，以学习图像特定的曲线映射，通过迭代的、逐像素的高阶曲线来增强低光图像。它使用非参考损失来引导学习，并以轻量模型实现了最先进的结果。

ABSTRACT

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.

研究动机与目标

• Motivate low-light enhancement without relying on paired or unpaired training data.
• Propose a data-efficient deep curve estimation framework to adjust dynamic range per image.
• Design differentiable non-reference losses to implicitly guide quality of enhancement.
• Develop a lightweight network that predicts per-pixel curve parameters for high-order adjustments.

提出的方法

• Formulate enhancement as image-specific curve estimation using a quadratic Light-Enhancement curve LE(I(x);α) = I(x) + α I(x)(1 - I(x)).
• Extend LE to higher orders by iterating LE to obtain LE_n, and make the per-pixel α a map A(x) for pixel-wise adjustments.
• Implement DCE-Net, a 7-layer CNN that outputs 24 parameter maps (for 8 iterations across 3 channels) with a final Tanh activation.
• Train with non-reference losses including spatial consistency, exposure control, color constancy, and illumination smoothness.
• Apply losses in a differentiable framework to enable end-to-end zero-reference learning.
• Maintain pixel value range, monotonicity, and differentiability to preserve image structure during mapping.

实验结果

研究问题

• RQ1 Can a low-light enhancement model be trained without any reference images (paired or unpaired)?
• RQ2 Can image-specific, per-pixel curve mappings robustly adjust dynamic range across diverse lighting conditions?
• RQ3 Do non-reference losses sufficiently guide high-quality enhancement without ground-truth images?
• RQ4 Is a lightweight DCE-Net capable of competitive performance and real-time inference?
• RQ5 Can the method improve downstream tasks such as face detection in the dark?

主要发现

• Zero-DCE achieves competitive or superior qualitative and quantitative results without any reference data.
• The 7-layer DCE-Net with 8 iterations and pixel-wise curve maps attains strong PSNR and SSIM while minimizing MAE on benchmark sets.
• Non-reference losses (spatial consistency, exposure control, color constancy, illumination smoothness) contribute significantly to image quality, as shown by ablations.
• Zero-DCE runs in real-time (about 500 FPS for 640x480x3 images on GPU) and trains in about 30 minutes.
• The method improves face detection under low light compared with several baselines.
• Zero-DCE demonstrates favorable runtime and generalization across varied lighting conditions.