Deep learning based damage detection of concrete structures

Abstract

Damage detection of any civil engineering structure is a part of the interest in the field of engineering from which one can estimate the stability and lifetime of the structure. With the advancement of technology in the field of infrastructure, damage assessment of any structure with the help of convolutional neural networks (CNNs) and deep learning is gaining importance because of the ease with which it can detect the damage. By using these computer-aided techniques, we can reduce manpower and detect damage in inaccessible places that we cannot see directly. In the current study, we used ResNet-50, which is a part of a deep convolutional neural network with 50 layers deep in it. ResNet-50 is a subclass of convolutional neural networks most popularly used for image classification. Furthermore, we used the image data set collected from the Utah State University, Logan, Utah, USA. This data set contains nearly 56,000 annotated images of both cracked and non-cracked bridge deck images, wall images, and pavement images. These images contain a variety of obstructions, such as shadows and surface roughness in some cases. The present study aimed to train and test the images and compare the accuracy of the results among themselves with an increase in training data. We devised an algorithm to measure cracks as soon as we detected them. The results show that the Resnet-50 architecture was in good agreement with the developed algorithm.