Automated Image-Based Crack Detection in Masonry Structures: A Novel Application of Convolutional Neural Networks

Abstract

We are surrounded by brick structures which, though generally resilient, can be sometimes fail. Masonry failures, whether a single falling brick or an entire structural collapse, are extremely dangerous. Frequent inspections are an important part of reducing risk of such events. Previous research has successfully been able to identify cracks in other materials. However, identifying cracks in masonry is more challenging and has not yet been done.

This paper is the first to demonstrate a viable method for image-based crack detection in masonry structures. This was done using machine learning techniques. A convolutional neural network was trained on images taken of brick walls built in the lab and was then able to achieve 86% accuracy in classifying images taken from actual brick structures as cracked or uncracked. The results from the CNN were compared to those from a simpler algorithm: a support vector machine. The SVM showed similar accuracy to the CNN in making predictions on lab data which looked similar to training data, but accuracy did not transfer as well to the real-world data.