Concrete Surface Crack Segmentation Based on Deep Learning

Abstract

Structural health monitoring becomes popular and important in the field of structural engineering because this technology can elongate the structural life cycle as well as protect structures from natural hazards. In the past, structural health monitoring mostly relied on the contact sensors to acquire structural responses and then diagnosed structures from these measurements. Therefore, this study presents a deep learning-based method which can detect and segment the concrete cracks through the noncontact measurements, e.g., images. This method implements deep learning with computer vision to identify crack existence and to further perform segmentation. First, the training data are prepared by collecting images of concrete surfaces with/without cracks, and two state-of-the-art models such as DeepLabv3+ and Mask R-CNN are established along with the transfer learning methods and trained by real crack images. Then, the trained models can extract crack features and yield a mask (i.e. probability map). The cracks are identified and segmented in images from the predicted mask. Finally, the pixel-wise result is processed to determine the geometric properties of cracks such as lengths and widths. Three experiments are designed to examine these two models, and performance of these two models are evaluated with the mean intersection-over-union (mIoU) ratings. Moreover, a comprehensive comparison between Mask R-CNN and DeepLabv3+ is carried out. To sum up, cracks on the concrete surface can be successfully identified, and the near optimal selection of segmentation models under different scenarios is discussed and provided in this study.