ASR crack identification in bridges using deep learning and texture analysis. (April 2023)

• Record Type: Journal Article
• Authors: Nguyen, Andy; Gharehbaghi, Vahidreza; Le, Ngoc Thach; Sterling, Lucinda; Chaudhry, Umar Inayat; Crawford, Shane
• Is Part Of: Structures. Volume 50(2023)
• Page Start: 494
• Page End: 507
• Publication Date: 2023-04
• Subjects: Alkali-Silica Reaction -- Structural Health Monitoring -- Artificial Intelligence -- Deep Learning -- Feature Enhancement -- Texture Analysis -- Two-stage Assessment; Structural engineering -- Periodicals; 624.1
• DOI: 10.1016/j.istruc.2023.02.042
• Languages: English
• ISSNs: 2352-0124

Abstract

Abstract: Alkali-Silica Reaction (ASR), commonly known as 'concrete cancer, ' is an expansive reaction occurring over time between aggregate constituents and alkaline hydroxides from cement. As a destructive phenomenon, the need to detect the onset of ASR in concrete structures to ensure their long-term durability and structural integrity is thus evidenced. In the structural health monitoring field, vision-based approaches have been found to be viable, fast, and cost-effective in diagnosing numerous types of cracks using physical attributes and surface patterns. However, achieving high accuracy in detecting ASR cracks using traditional visual inspection techniques is challenging and time-consuming. Inspired by artificial intelligence technology, this paper proposes and evaluates a two-phase computer vision procedure for detecting and classifying ASR cracks utilizing a collection of ASR images recorded from several bridges in Queensland, Australia. In the first phase, the procedure compares common pre-trained CNN models to investigate their capability in classifying ASR cracks and to select the best-performed model. In the second phase, a novel Feature Enhancement Process (FEP) was first proposed to increase the contrast between ASR cracks and the heavily textured backgrounds within the images. Next, to better highlight the ASR crack features, the feature-adjusted images are processed further through different texture analysis algorithms including: (i) Texture Morphology, (ii) Adaptive thresholding, and (iii) Local range filtering. The study shows that the proposed FEP can improve the ASR crack classification accuracy of InceptionV3, which is the best CNN model selected from Phase 1, from 90.9% to 92.48%. Furthermore, by combining FEP with texture morphology, a robust two-stage tool for assessing ASR cracks can be made with an impressive validation accuracy of 94.07%. This research contributes towards the application of novel AI deep learning technology in providing cost-effective autonomous ASR crack classification tools to support the owners and managers of civil public works assets and other constructed infrastructures.