Multistage semisupervised active learning framework for crack identification, segmentation, and measurement of bridges

Abstract In bridge health monitoring (BHM), crack identification and width measurement are two of the most important indices for evaluating the functionality of bridges. In order to reduce the labor cost in field detection, researchers have proposed a variety of deep learning (DL)‐based detection techniques for crack recognition. However, some problems still exist in extending these techniques to practical applications, such as data annotation difficulty, limited model generalization ability, and inaccuracy of the DL identification of the actual crack width measurement. In this paper, an application‐oriented multistage crack recognition framework is proposed, namely, C onvolutional A ctive L earning I dentification‐ S egmentation‐ M easurement (CAL‐ISM). It includes four steps: (1) pretraining of the benchmark classification model, (2) retraining of the semisupervised active learning model, (3) pixel‐level crack segmentation, and (4) crack width measurement. Beyond numerical experiments, the performance of the CAL‐ISM is validated for practical applications: (i) bridge column test specimen and (ii) field BHM projects. In conclusion, the obtained results from these applications shed light on the high potential of CAL‐ISM for BHM applications, which is recommended in future deployments for BHM.