Crack Detection an‎d Identification in Concrete Structures Using Machine Learning Algorithms an‎d Image Processing

12/24/2026 12:00:00 AM

Ensuring the structural integrity of civil infrastructure is essential for reducing economic losses an‎d safeguarding human lives against natural hazards such as earthquakes. Among various indicators of structural deterioration, cracks in concrete members are recognized as the most prominent manifestations of damage. Therefore, accurate detection an‎d quantitative eva‎luation of cracks play a critical role in structural health monitoring (SHM). In this study, an intelligent deep learning–based framework was developed for automatic crack detection, segmentation, an‎d geometric analysis on concrete surfaces. In the first stage, six advanced segmentation architectures, including EfficientCrackNet, MSP U-Net, CPCDNet, PCTC-Net, Segment Any Crack (SAC), an‎d DeepLabV3+, were comparatively eva‎luated. The results indicated that DeepLabV3+ achieved superior performance an‎d was therefore selec‎ted as the base framework for further optimization. In the second stage, six backbone networks—ResNet-50, ResNet-101, MobileNetV2, GoogleNet, AlexNet, an‎d VGG16—were investigated to determine the optimal feature extractor. Among them, ResNet-101 demonstrated the best overall performance, achieving an accuracy of 98.28% an‎d a test error of 0.0426, outperforming other configurations in terms of training stability an‎d convergence behavior. The proposed model was trained using transfer learning with pre-trained weights from the COCO dataset for binary classification (crack/non-crack). A dataset consisting of 20,000 image–mask pairs was prepared an‎d divided into training (14,000), validation (3,000), an‎d testing (3,000) subsets. Adaptive thresholding was applied as a post-processing step to convert network outputs into precise binary crack maps. For geometric feature extraction, several advanced image processing techniques were employed. Crack length was computed using a hybrid approach incorporating skeletonization, geodesic path optimization, an‎d cubic B-spline interpolation, achieving an accuracy of 95.5%. Crack width was measured through a combination of distance transform, sub-pixel estimation, an‎d Gaussian process regression, reaching mean an‎d maximum accuracies of 96.02% an‎d 95.97%, respectively. The slope of crack width variation was obtained via spatial differentiation an‎d polynomial fitting, while crack depth was estimated based on empirical correlations an‎d linear elastic fracture mechanics principles. Finally, cracks were classified into five severity levels—negligible, low, moderate, high, an‎d very high—with an overall accuracy of 95.33%. The integration of multi-scale feature extraction with precise geometric analysis provides a robust an‎d efficient tool for structural health monitoring applications. The findings confirm that combining deep learning with advanced image processing algorithms can offer a reliable, accurate, an‎d cost-effective alternative to traditional manual inspection methods in civil engineering.

بخش‌بندي ترك , ارزيابي آسيب بتن , يادگيري انتقالي , پايش سلامت سازه , يادگيري عميق

Crack Segmentation , Concrete Damage Assessment , Transfer Learning , Structural Health Monitoring (SHM) , Deep Learning

Seyed Tohid Aghadavood Jolfaie

Dr. Gholamreza Ghodrati Amiri