Abstract: To address the difficulty of rock fracture detection in the water-level fluctuation zone of the Three Gorges Reservoir under low contrast and complex background conditions, as well as the large number of parameters and high computational cost of the UNet model that limit its deployment on edge devices, a lightweight semantic segmentation model, LiteUNet, is proposed. Based on the UNet architecture, depthwise separable convolution, a multi-scale context module, a lightweight decoder, and a deep supervision mechanism are introduced to improve the detection of fine fractures and complex boundaries. A dataset of rock fracture images from the Wushan section of the fluctuation zone was constructed, and data augmentation was applied to expand the sample size. The proposed model was compared with several mainstream semantic segmentation models under the same experimental conditions. The results show that LiteUNet achieves an mIoU of 82.92% and an F1-score of 90.66%. The number of parameters and computational cost are reduced to 2.76 M and 17.61 GFLOPs, respectively, which are significantly lower than those of mainstream models. In practical applications, fracture information in the fluctuation zone can be stably extracted, and a good balance between accuracy and efficiency is achieved. Therefore, the proposed model is suitable for fast fracture detection and deployment on edge devices in complex fluctuation zone scenarios.