Abstract: Traditional manual inspection of water leakage in tunnel linings is highly subjective and inefficient, while existing deep learning models still have limited recognition accuracy in complex scenarios. To improve the accuracy and efficiency of water-leakage detection in tunnel linings, this study develops an intelligent water-leakage recognition model for complex backgrounds. A Vision Transformer (ViT) was incorporated into the UNet architecture to form a TransUNet network. A squeeze-and-excitation block (SE-Block) channel attention mechanism was introduced to weight global semantic features, and the skip connections between the encoder and decoder were optimized, thereby constructing the SE-TransUNet model. The model was trained on a mixed water-leakage dataset. Ablation experiments were conducted to verify the effectiveness of the SE-Block and ViT modules, and score-weighted class activation mapping (Score-CAM) heatmaps were used to analyze the model’s attention mechanism for water-leakage features. The results show that: (1) compared with eight mainstream semantic segmentation models, SE-TransUNet achieves better overall performance, with intersection over union (IoU), recall, precision, accuracy, and F1-score values of 0.8325, 0.9541, 0.8674, 0.9487, and 0.9082, respectively, and shows strong adaptability to scenarios involving indentation shadows, stain interference, and faint water-leakage traces; (2) the ablation results confirm that introducing the SE and ViT modules has a positive synergistic effect on model performance, with the IoU of the baseline model being 0.0461 and 0.0711 higher than those of the variants without all SE modules and without both SE and ViT modules, respectively; and (3) the Score-CAM heatmaps and deletion/insertion curves indicate that the SE module helps expand the model’s attention coverage over water-leakage regions, enhance feature continuity, and improve resistance to interference in complex environments. The proposed SE-TransUNet model demonstrates effective water-leakage recognition capability with the dataset and experimental settings used in this study, providing a reference for intelligent detection of tunnel lining defects.