Automatic crack classification and segmentation on concrete bridge images using convolutional neural networks and hybrid image processing

Abstract Cracks are an indicator for a bridge’s structural health and functional failures. Crack detection is one of the major tasks needed to maintain the structural health and serviceability of a bridge. At present, the most commonly used crack detection technology is manual inspection, which has the disadvantages of being highly labor-intensive and time-consuming. In this paper, a crack detection method based on a convolutional neural network (CNN) is proposed. To automate quantitative measurements of an identified crack, hybrid image processing is proposed, as well. First, a dataset is compiled, including 12,000 cropped crack images and 19,500 cropped background images. Second, preprocessed images with the proposed method of Bilateral-Graying-Contrast (BGC) are fed into ResNet and a Visual Geometry Group Network (VGG) for training and testing. Finally, an automatic measurement system for bridge crack is developed which is not prone to weakened shooting conditions. The results demonstrate that ResNet achieves an accuracy of crack detection up to 97.44%, which is higher than VGG. Our crack measurement system significantly reduces the measurement error to 9.86% and can be assumed as a reliable method in the analysis of concrete bridge images.