This paper presents a deep Retinex decomposition method for low-light image enhancement. The method, called Retinex-Net, integrates image decomposition and subsequent enhancement operations. It consists of two main components: Decom-Net for decomposition and Enhance-Net for illumination adjustment. Decom-Net is trained on a large-scale dataset of low/normal-light image pairs, using constraints such as consistent reflectance and smooth illumination. Enhance-Net adjusts the illumination map while maintaining global consistency and local detail. The method also includes denoising operations on reflectance to handle amplified noise in dark regions. Extensive experiments demonstrate that the proposed method achieves visually pleasing low-light enhancement and provides a good representation of image decomposition. The contributions include the construction of a large-scale dataset, a deep-learning-based image decomposition method, and a structure-aware total variation constraint for smooth illumination maps.This paper presents a deep Retinex decomposition method for low-light image enhancement. The method, called Retinex-Net, integrates image decomposition and subsequent enhancement operations. It consists of two main components: Decom-Net for decomposition and Enhance-Net for illumination adjustment. Decom-Net is trained on a large-scale dataset of low/normal-light image pairs, using constraints such as consistent reflectance and smooth illumination. Enhance-Net adjusts the illumination map while maintaining global consistency and local detail. The method also includes denoising operations on reflectance to handle amplified noise in dark regions. Extensive experiments demonstrate that the proposed method achieves visually pleasing low-light enhancement and provides a good representation of image decomposition. The contributions include the construction of a large-scale dataset, a deep-learning-based image decomposition method, and a structure-aware total variation constraint for smooth illumination maps.