This paper presents a novel method called Zero-Reference Deep Curve Estimation (Zero-DCE) for low-light image enhancement. Zero-DCE formulates light enhancement as a task of image-specific curve estimation using a deep network. The 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 to consider pixel value range, monotonicity, and differentiability. Zero-DCE is appealing due to its relaxed assumption on reference images, as it does not require any paired or unpaired data during training. This is achieved through a set of carefully formulated non-reference loss functions that implicitly measure the enhancement quality and drive the learning of the network. The method is efficient as image enhancement can be achieved by an intuitive and simple nonlinear curve mapping. Despite its simplicity, it generalizes well to diverse lighting conditions. Extensive experiments on various benchmarks demonstrate the advantages of the method over state-of-the-art methods both qualitatively and quantitatively. Furthermore, the potential benefits of Zero-DCE to face detection in the dark are discussed. The proposed method is lightweight, capable of processing images in real-time, and takes only 30 minutes for training. It is superior to existing data-driven methods in three aspects: it explores a new learning strategy requiring zero reference, it uses carefully defined non-reference loss functions, and it is highly efficient and cost-effective. The method is capable of improving high-level visual tasks, such as face detection, without inflicting high computational burden. The framework of Zero-DCE is presented, including the Light-Enhancement Curve (LE-curve), DCE-Net, and non-reference loss functions. The LE-curve is designed to map a low-light image to its enhanced version automatically, with self-adaptive curve parameters dependent on the input image. The DCE-Net is a deep curve estimation network that estimates a set of best-fitting Light-Enhancement curves given an input image. The non-reference loss functions include spatial consistency loss, exposure control loss, color constancy loss, and illumination smoothness loss. The method is evaluated on various benchmarks and shows superior performance in terms of visual quality and perceptual quality. It is also effective for face detection in low-light conditions. The proposed method is efficient, lightweight, and capable of processing images in real-time.This paper presents a novel method called Zero-Reference Deep Curve Estimation (Zero-DCE) for low-light image enhancement. Zero-DCE formulates light enhancement as a task of image-specific curve estimation using a deep network. The 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 to consider pixel value range, monotonicity, and differentiability. Zero-DCE is appealing due to its relaxed assumption on reference images, as it does not require any paired or unpaired data during training. This is achieved through a set of carefully formulated non-reference loss functions that implicitly measure the enhancement quality and drive the learning of the network. The method is efficient as image enhancement can be achieved by an intuitive and simple nonlinear curve mapping. Despite its simplicity, it generalizes well to diverse lighting conditions. Extensive experiments on various benchmarks demonstrate the advantages of the method over state-of-the-art methods both qualitatively and quantitatively. Furthermore, the potential benefits of Zero-DCE to face detection in the dark are discussed. The proposed method is lightweight, capable of processing images in real-time, and takes only 30 minutes for training. It is superior to existing data-driven methods in three aspects: it explores a new learning strategy requiring zero reference, it uses carefully defined non-reference loss functions, and it is highly efficient and cost-effective. The method is capable of improving high-level visual tasks, such as face detection, without inflicting high computational burden. The framework of Zero-DCE is presented, including the Light-Enhancement Curve (LE-curve), DCE-Net, and non-reference loss functions. The LE-curve is designed to map a low-light image to its enhanced version automatically, with self-adaptive curve parameters dependent on the input image. The DCE-Net is a deep curve estimation network that estimates a set of best-fitting Light-Enhancement curves given an input image. The non-reference loss functions include spatial consistency loss, exposure control loss, color constancy loss, and illumination smoothness loss. The method is evaluated on various benchmarks and shows superior performance in terms of visual quality and perceptual quality. It is also effective for face detection in low-light conditions. The proposed method is efficient, lightweight, and capable of processing images in real-time.