Machine Vision Method for Crack Detection and Pattern Recognition in UHPC Prestressed Beams

Abstract

Although ultra-high-performance concrete (UHPC) is known for its exceptional durability and crack resistance, UHPC bridges in harsh environments can still develop cracks. Detecting cracks in concrete structures, analyzing crack patterns, and determining the types of cracks that lead to structural failure have always posed significant challenges for the industry. Cracks in UHPC structures are typically finer and more densely distributed compared to those in normal concrete. Manual visual observation is susceptible to subjective errors and often requires considerable time and professional expertise, which hinders swift postdisaster rescue efforts. Machine vision technology offers a promising solution for visualizing cracks and identifying faults in concrete structures. This study conducted material and structural experiments on a UHPC prestressed beam, varying the shear span ratio, stirrup ratio, and external prestress arrangement. Through an analysis of the crack initiation trend, development process, and failure mode of the test beam, a method is proposed to achieve higher precision in extracting fine cracks using image preprocessing techniques. Subsequently, a new stacking-algorithm automatic crack classifier was used to identify crack fault patterns. Its performance was compared with five traditional machine learning classifiers based on accuracy, precision, recall, F1 score, and confusion matrix. The results demonstrated that among the six classifiers, the proposed stacking classifier for automatic crack inspection achieved the highest accuracy at 98.33%. Finally, integrating these technologies into a smartphone terminal enables convenient offline detection and classification of cracks in UHPC beams.