Cost-effective system for detection and quantification of concrete surface cracks by combination of convolutional neural network and image processing techniques

Abstract

This study proposed a semi-automated system for crack detection and quantification, based on the combination of a trained convolutional neural network (CNN) and a developed application. Specifically, we tested four commonly used CNNs and determined GoogLeNet for this study. Then, the transfer learning and fully training of GoogLeNet were further tested on our testing dataset and a public dataset. The results show that the transfer learning GoogLeNet has relatively balanced performances on these two datasets, with accuracy of 96.69% and 88.39%, respectively. A new sliding window technique (neighborhood scanning) was proposed and shown almost equivalent performance to the previous dual scanning method. A method for calculating crack width was presented. The average relative error of this method is 14.58% (0.05 mm), i.e., much smaller than the 36.37% (i.e., 0.14 mm) of the previous method. An application was then developed to integrate the proposed methods and other techniques such as edge detectors, boundary tracking, and threshold segmentation to segment, quantify, and analyze cracks. Verifications on 23 untrained raw images (eleven with 10240 × 2048 pixels, twelve with 2592 × 4608 pixels) show that: (1) the developed system and a previous pixel-level segmentation system require an average of 9.48 s and 10.35 s; (2) these two systems show an 80.40% and a 78.64% average Intersection over Union (IoU). Therefore, the proposed system is a cost-effective solution for detecting and analyzing cracks on concrete surfaces considering its practical performance and time cost. Practically, the proposed system could be used to analyze the images collected from onsite inspection or from experiment.