Toward Automation in Crack Detection and Measurements: Benchmarking of CNN-based Algorithms

Cracks are one of the main defect features on concrete surfaces, and indicators of concrete structures' condition regarding their state of health. Since traditional methods to identify and assess cracks rely on manual measurements, a significant number of studies to date have focused on identifying ways to automate this process. Accordingly, this study proposes to combine Convolutional Neural Network (CNN)-based algorithms and traditional image morphological operation for crack detection and measurements with a specific goal to benchmark the proposed methodology on data of three images obtained from laboratory experiments in a controlled environment. The proposed methodology performed well, with 92.10% and 90.11% accuracy in crack length and width measurements, respectively. Future research will focus on fine-tuning the proposed crack detection and measurement methodology and evaluate them on a set of images acquired from a real full-scale structure such as a bridge or a building.