Image‐based concrete crack assessment using mask and region‐based convolutional neural network

Summary Recently, many countries have investigated replacing conventional visual inspection with computer‐vision–based inspection to enhance the efficiency, speed, and objectivity of inspection. This paper presents a novel crack assessment framework for concrete structures that detects cracks using mask and region‐based convolutional neural network (Mask R‐CNN) and quantifies cracks using a few morphological operations on the detected crack masks. In this study, a Mask R‐CNN is trained for crack detection using 1,102 crack regions masked on 376 concrete images. The trained Mask R‐CNN model is tested on the images taken from a real concrete wall with 453 cracks whose widths range from less than 0.1 mm to 1.0 mm. The trained model successfully detects most of the cracks 0.3 mm or wider. Quantification of the cracks was then carried out using several image‐processing operations on 10 randomly selected crack masks. Cracks with widths of 0.3 mm or more are quantified successfully with errors less than 0.1 mm, whereas cracks less than 0.3 mm widths show relatively larger error due to the limitation of image resolution.