Zhao, W.; Li, K.; Liu, G.; Chen, B.; Fan, M. and Yin, S. (2026) A three-stage intelligent crack detection method for concrete bridges based on deep neural networks. Journal of Construction Engineering and Management, 152(6): 04026067, ISSN 0733-9364
Abstract
Crack detection is a critical task in the diagnosis and assessment of bridge health. Currently, most crack detection methods concentrate solely on a single aspect (segmentation, object detection, or classification) and are incapable of fulfilling the engineering requirements across various stages of real-world bridge crack detection projects, which is insufficient to provide comprehensive information for subsequent crack diagnosis and maintenance. This study proposes a three-stage intelligent crack (THICK) detection method for concrete bridges based on deep neural networks, which includes the Crack-Filter neural network for filtering out pseudocrack images, the Transformer Residual U-net neural network for pixel-level segmentation of cracks, and the Crack-Class neural network for the classification of bridge cracks. By sequentially passing the images to be detected through these three neural networks, it can not only realize the individual functions of filtering, localization, and classification, but also improve the detection accuracy of the latter function through the former function. The method is tested on three self-constructed and several authoritative data sets, demonstrating its strong generalization performance and engineering capabilities. Specifically, in the test data set of images captured from actual projects, after applying Crack-Filter to eliminate pseudocracks, the mean intersection over union (mIOU) using Tres-net to segment reaches 0.6652, representing a significant improvement of 45% compared to the value obtained before processing (0.46). Moreover, following the combined processing of Crack-Filter and Tres-net joint treatment and subsequently utilizing Crack-Class stage for crack classification, the F1 score of crack classification can be enhanced to 0.93, marking a 52% improvement compared to the score obtained before joint treatment (0.61). In summary, the THICK detection method offers a comprehensive, automated, and intelligent solution for detecting bridge cracks and enables significant enhancement of segmentation and classification accuracy.
| Item Type: | Article |
|---|---|
| Uncontrolled Keywords: | crack classification; crack detection; deep neural networks; fully automated; semantic segmentation |
| Index terms: | object detection, neural network, localization, concrete bridge, accuracy |
| Subjects: | urban planning, artificial intelligence, professional development, computer vision, infrastructure and transport systems |
| Topics: | Engineering Principles, Digital Applications, Governance, Information Management |
| Descriptive scope: | 2 PC |
N.B. Descriptive scope is a count of how many of the five facets of empirical research are indicated by the words used in title, abstract and keywords. It is not intended as a judgement on the research; merely a count of the kind of word we would expect to indicate Phenomenon, Concepts, Theoretical framing, Empirical techniques, Analytical techniques. If all five are present, then a code of “5 PCTEA” will indicate this. If you feel the coding for this record is questionable, we welcome discussion around the terms we matched or the way we categorized them. The facet you would expect may not be coded, or a facet may be coded inappropriately. This can also bear on a larger question, of which facets should be treated as defining in construction management research. Please get in touch, and we will look at it. More details here