In the field of computer vision, enhancing low-light images is a significant challenge, primarily due to the reliance on high-quality paired low-light and high-light images in supervised learning methods, which are expensive to acquire. This paper presents an unpaired approach for low-light image enhancement, integrating a Joint Estimation Network and a Multi-Domain Feature Fusion Network. The Joint Estimation Network is trained exclusively with pairs of low-light images of the same scene, while the Multi-Domain Feature Fusion Network is trained solely with normal-light images that are not paired with the aforementioned low-light images. The Joint Estimation Network decomposes low-light images into components of illumination, reflectance, and noise. After enhancing the illumination, it passes these components, along with the reflectance, to the Multi-Domain Feature Fusion Network. The Multi-Domain Feature Fusion Network employs multi-scale encoder-decoder modules and frequency domain adjustments to enhance details and maintain global consistency. Our method addresses the issues of insufficient illumination and high noise in low-light images, improving visual quality without the need for paired images, thereby increasing the model’s practicality in real-world applications.