Automated Defect Detection For Masonry Arch Bridges

Abstract

The condition of masonry arch bridges is predominantly monitored with manual visual inspection. This process has been found to be subjective, relying on an inspection engineer’s interpretation of the condition of the structure. This paper initially presents a workflow that has been developed that can be used by a future automated bridge monitoring system to determine underlying faults in a bridge and suggest appropriate remedial action based on a set of detectable symptoms. This workflow has been used to identify the main classes of defects that an automated visual detection system for masonry should be capable of detecting. Subsequently, a convolutional neural network is used to classify these identified defect classes from images of masonry. As the mortar joints in the masonry are more distinctive than the defects being sought, their effect on the performance of an automated defect classifier is investigated. Compared to classifying all the regions of the masonry with a single classifier, it is found that where the mortar and brick regions have been classified separately, defect and defect free areas of the masonry have been predicted both with more confidence and with better accuracy.