Abstract: In low-light coal mine underground scenarios, captured images typically exhibit issues, such as insufficient brightness, low contrast, and noise interference, severely degrading image quality and hindering the accuracy of subsequent visual processing tasks. Traditional low-light image enhancement algorithms are prone to show problems like image distortion, over-exposure, and low contrast in complex environments, unsupervised learning methods lack precise guidance signal, leading to suboptimal detail recovery and color restoration compared to supervised approaches. A novel supervised decoupled cascade denoising network is proposed to address this. The network operates in the HVI color space, effectively separating the image illumination intensity and chrominance, and applying linear transformation to the chrominance component to avoid the suppression of negative-value features by nonlinear activation functions. Additionally, the perturbations encountered during low-light image enhancement are decoupled into two types: Overexposure and underexposure disturbances caused by illumination estimation errors that affect the image’s brightness uniformity, and blurring and artifact disturbances caused by noise in the reflectance component that impact the image’s color accuracy based on the improved Retinex theory. And two U-Net specialized denoisers are respectively designed to eliminate the noise in two decoupled branches. Specifically, denoiser L is employed to approximate the inverse matrix of the illumination branch disturbance, while denoiser R is used to approximate the negative matrix of the reflectance branch disturbance, thereby enhancing the natural brightness and color fidelity of low-light images. Furthermore, an efficient Linear Transformer model with KAN network is designed as the core component of the denoisers for boosting performance. By introducing group-wise parameter sharing strategy and rational polynomial functions to replace the conventional MLP layers in the Transformer architecture, this network possesses strong nonlinear fitting capabilities while significantly reducing the parameters, greatly boosting the expressiveness and computational efficiency. Experiments demonstrate that the proposed algorithm achieves SOTA performance on both the public LOL dataset and the coal mine underground dataset, exhibiting remarkable generalization capability in Zero-shot testing tasks within coal mine underground scenes.