DWCF-YOLO: tunnel leakage detection and segmentation based on multi-level wavelet transform convolution

Abstract Subway tunnels are essential components of urban transportation systems, and water leakage can threaten their structural safety. Accurate detection and segmentation of leakage regions are therefore critical. This study develops a tunnel water leakage dataset from point-cloud data collected using a mobile measurement platform and converts the point clouds into grayscale images. We propose DWCF-YOLO, an enhanced YOLOv11-based network for water leakage detection and segmentation. The model incorporates multi-level wavelet transform convolution to extract multi-frequency features and promote local–global feature fusion, improving the identification of leakage areas at different scales. In addition, content-guided attention provides channel- and spatial-directional guidance for multi-scale feature maps, enhancing sensitivity to leakage regions. Experimental results show that DWCF-YOLO achieves 96.8% accuracy, 97.8% recall, 98.4% F1-score, 98.9% precision, 96.8% mIoU, and 97.8% mPA. Compared with mainstream models, it obtains the best performance in four of the six evaluation metrics, demonstrating its effectiveness and reliability for tunnel water leakage detection.