Deep Learning-Based Method for Detecting Defects Under Urban Utility Tunnels

Urban Utility Tunnel (UUT) is susceptible to defects throughout its service life due to various factors, such as material aging, soil load, and adverse conditions. These defects deteriorate the tunnel's load-bearing capability and compromise the integrity of pipeline facilities. Hence, there is an urgent need for an effective method to detect UUT defects. In this study, we propose an Enhanced U-Net model called EU-Net, which leverages ResNet34 as the feature extraction backbone in place of the base U-Net. This substitution significantly enhances both global and local feature extraction capabilities. Additionally, we introduce a spatial attention module to establish meaningful relationships between defect pixels at different distances, thereby reinforcing important information. To tackle the challenge of limited samples and prevent overfitting, we employ a data augmentation method to increase dataset diversity. Compared to the conventional U-Net model, our improved U-Net achieves a notable 13.7% increase in Fl-score, a 10.1% increase in Mean Intersection of Union (MIoU), faster detection speed, and improved accuracy. These results convincingly demonstrate the effectiveness and practicality of our proposed method in real-world UUT defect identification tasks.