This paper proposes a deep Retinex-Net for low-light image enhancement, which decomposes images into reflectance and illumination components and adjusts illumination for lightness enhancement. The method is trained on a new dataset, LOL, containing 500 low/normal-light image pairs. The network consists of two subnetworks: Decom-Net for decomposition and Enhance-Net for illumination adjustment. Decom-Net learns to decompose images without ground truth, using constraints such as shared reflectance between low/normal-light images and smooth illumination. Enhance-Net adjusts illumination to maintain consistency while tailoring local distributions. A structure-aware total variation loss is introduced to ensure smooth illumination while preserving structures. Denoising is applied to reflectance to handle noise in dark regions. The network is end-to-end trained, allowing for natural light adjustment. Extensive experiments show that the method achieves visually pleasing results and provides good image decomposition. The dataset is the first of its kind for low-light enhancement, and the method outperforms existing techniques in low-light enhancement and denoising. The approach is data-driven, leveraging deep learning to improve the decomposition and enhancement process.This paper proposes a deep Retinex-Net for low-light image enhancement, which decomposes images into reflectance and illumination components and adjusts illumination for lightness enhancement. The method is trained on a new dataset, LOL, containing 500 low/normal-light image pairs. The network consists of two subnetworks: Decom-Net for decomposition and Enhance-Net for illumination adjustment. Decom-Net learns to decompose images without ground truth, using constraints such as shared reflectance between low/normal-light images and smooth illumination. Enhance-Net adjusts illumination to maintain consistency while tailoring local distributions. A structure-aware total variation loss is introduced to ensure smooth illumination while preserving structures. Denoising is applied to reflectance to handle noise in dark regions. The network is end-to-end trained, allowing for natural light adjustment. Extensive experiments show that the method achieves visually pleasing results and provides good image decomposition. The dataset is the first of its kind for low-light enhancement, and the method outperforms existing techniques in low-light enhancement and denoising. The approach is data-driven, leveraging deep learning to improve the decomposition and enhancement process.