Concrete Crack Detection and Classification Methods Based on Machine Vision and Deep Learning

With the rapid development of underground space, structural crack monitoring has become increasingly critical. This study proposes a unified framework integrating image preprocessing, feature extraction, model training, and safety assessment for crack analysis. An improved OTSU threshold segmentation algorithm based on sliding windows and local statistical analysis is developed to enhance noise suppression and detail preservation under complex backgrounds and varying resolutions. For crack identification and orientation classification, SVM, CNN, ResNet-18, and K-means clustering are systematically compared. The results show that the improved OTSU method outperforms the classical approach in both high- and low-resolution images. In classification tasks, SVM achieves the best performance under limited data conditions, with accuracy exceeding 96% and reaching 97% after outlier removal, outperforming CNN, K-means, and ResNet-18. Although ResNet-18 demonstrates strong overall performance with high prediction confidence across crack categories, it remains slightly inferior to SVM when training data are limited. Experimental validation using full-scale loading tests of metro shield tunnel segments further confirms the robustness of the proposed approach, with SVM achieving an accuracy of 95.45% in real-world conditions. This study provides an efficient and reliable solution for automated crack detection and classification in metro tunnel infrastructure and similar underground segment-based systems.