Abstract: Computer vision technology offers an intelligent solution for tunnel seepage area detection, yet its feature extraction and segmentation accuracy remains constrained by low-light conditions and pipeline obstructions. This study proposes a Light-UNet model based on the Unet model, an enhanced tunnel seepage zone segmentation network with three key contributions: (1) To enhance the extraction accuracy of the seepage area, a light-shadow transformation boot item is introduced, and a parallel attention module is designed to extract internal information of the boot item. Subsequently, channel attention mechanisms are employed to aggregate the boot item with deep network information. (2) To overcome the scarcity of tunnel seepage datasets, the data was manually collected from a tunnel under construction. Perspective transformation was used to alter camera angles and expand the dataset. (3) To address insufficient lighting in tunnels, Gaussian noise and salt-and-pepper noise were randomly added to the dataset images to simulate images captured in dim environments, thereby enhancing the network’s generalization ability. Ultimately, the Light-Unet network was trained using the constructed training set, achieving a mean Intersection over Union (mIoU) of 88.84% and a mean Dice coefficient (mDice) of 88.04%, both significantly higher than the Unet baseline model. Visualization comparisons between the improved and baseline networks demonstrate that the improved network better adapts to the segmentation of seepage areas in complex environments, including poor lighting and pipeline obstructions.