Automated classification of civil structures defects based on Convolutional Neural Network

Today, the most used method for civil infrastructure inspection is based on visual assessment performed by certified inspectors following prescribed protocols. However, the increase in aggressive environmental and load conditions, coupled with the achievement for many structures of the end life-cycle, highlighted the need to automate damage identification to satisfy the number of structures that need to be inspected. To overcome this challenge, the current paper presents a method to automate the concrete damage classification using a deep Convolutional Neural Network (CNN). The CNN is designed after an experimental investigation among a wide number of pretrained networks, all applying the transfer learning technique. Training and Validation are performed using a built database with 1352 images balanced between “undamaged”, “cracked”, and “delaminated” concrete surface. To increase the network robustness compared to images with real-world situations, different configurations of images has been collected from Internet and on-field bridge inspections. The GoogLeNet model is selected as the most suitable network for the concrete damage classification, having the highest validation accuracy of about 94%. The results confirm that the proposed model can correctly classify images from real concrete surface of bridges, tunnel and pavement, resulting an effective alternative to the current visual inspection.