Automated bughole detection and quality performance assessment of concrete using image processing and deep convolutional neural networks. (26th April 2021)

• Record Type: Journal Article
• Authors: Wei, Wei; Ding, Lieyun; Luo, Hanbin; Li, Chen; Li, Guowei
• Is Part Of: Construction & building materials. Volume 281(2021)
• Publication Date: 2021-04-26
• Subjects: Concrete -- Bughole -- Image processing -- Deep convolutional neural network -- Quality performance assessment -- Structural health monitoring; Building materials -- Periodicals; 624.18
• DOI: 10.1016/j.conbuildmat.2021.122576
• Languages: English
• ISSNs: 0950-0618

Abstract

Highlights: An innovative method is proposed for the detection of concrete bugholes. Combined use of image processing and DCNN provides a quick and reliable detection. An inception module consisting of multiple operations was introduced into the DCNN. Quality performance of concrete was assessed based on the detection results. The potential relationship between the attributes of bugholes and compressive strength of concrete was revealed. Abstract: Concrete is ubiquitous in the modern construction and building industry, bugholes are one of the most typical imperfections on concrete surface, and traditional manual inspection techniques are inefficient and difficult to rapidly deploy. With the development of image processing and deep learning techniques, new opportunities and possible solutions are available for bughole detection. This paper not only dedicates to propose a novel approach to automatically extract bugholes in the concrete surface images with high fidelity by combining the advantages of image processing and deep convolutional neural network (DCNN), but also attempts to assess the quality performance of concrete based on the detection results. Firstly, a series of image processing procedures were performed to highlight the characteristics of bugholes in the acquired images and reduce the data size. Secondly, a DCNN was proposed to automatically detect bugholes in the images, an inception module consisting of multiple convolutions and a max pooling operation was introduced into the network to efficiently extract multiscale features and reduce the model parameters. Favorable detection results were obtained in the experimental tests, especially the small bugholes (i.e., diameter ranging from 0.3 mm to 1 mm) can be accurately detected. Finally, the quality performance of concrete specimens was assessed based on the detection results, the potential relationship between the attributes of bugholes and compressive strength of concrete specimens was revealed through a fitting function. The proposed method makes both the detection of bugholes and assessment of concrete automatable, efficient, and low cost, potentially aiding engineers in inspecting concrete structures in a fast and reliable manner.