To enhance the efficiency and accuracy of tunnel leakage detection, reduce false and missed detection rates, and mitigate the impact of subjective judgment and varying expertise in manual detection methods, a novel segmentation approach based on an improved U-Net model was proposed. This method addressed the challenges posed by low-light conditions within tunnels, which negatively affected both imaging quality and segmentation performance. The EfficientNet-B1 network, developed by Google, was used as the encoder backbone for the U-Net model. Furthermore, a gated mechanism was employed to replace the traditional concatenation operation in the U-Net skip connections, and an squeeze-and-excitation (SE) module was incorporated into the gating mechanism. Comparative experiments were performed between the improved model, the original U-Net, and FCN models. The results show that the original U-Net achieves a mean intersection over union(mIoU) of 75.08%, an average Dice coefficient of 0.847, and an average Hausdor distance of 5.01. By replacing the skip connection concatenation with the SE-enhanced gating mechanism, the mIoU is increased to 78.92%, the average Dice coefficient improves to 0.877, and the Hausdor distance decreases to 4.93. The integration of EfficientNet-B1 as the U-Net encoder further boosts the mIoU to 84.24%, the average Dice coefficient to 0.914, and the Hausdor distance is reduced to 3.792, achieving the best performance among all models. Overall, a 9.16% improvement in mIoU is achieved compared to the original U-Net, and a 4.32% improvement is noted compared to the fully convolutional networks(FCN) model. These results demonstrate that the proposed method not only enhances the model's segmentation and recognition capabilities but also significantly improves its performance in identifying infrared leakage images.