Deep Neural Networks for Visual Bridge Inspections and Defect Visualisation in Civil Engineering
Year: 2021
Type: conference-paper
Abstract: Ageing infrastructure is a global concern, and current structural health monitoring practices are coming under review. With a view to streamline the visual bridge inspection process, we assess the classification performance of two Deep Neural Networks, VGG16 and MobileNet, on a challenging dataset of over 70,000 unprocessed bridge inspection images of three defect categories: corrosion, crack, and spalling. Grad-CAM "heatmap" visualisations on VGG16 predictions provide a coarse localisation of the defect region and some insight into the functioning of the network. Similar performance is attained on MobileNet, for applications where speed or computational cost is a consideration. We conclude that with further optimisation this approach could have an application in automated defect tagging.
Authors: Julia Bush, Tadeo Corradi, Jelena Ninić, Georgia E. Thermou, John Bennetts
Language: English
Subfield: Civil and Structural Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W3174189849