Using machine learning approaches to perform defect detection of existing bridges

The paper presents a study about defect detection on structural elements of existing bridges through a machine-learning approach. In detail, the proposed methodology aims to explore the possibility of automatically recognizing defects and damages on bridges' elements, (e.g., cracks, humidity) by employing a training of existing convolutional neural networks on a set of photos. The initial database has been firstly selected and then classified by domain experts according to the requirements of the new Italian Guidelines on structural safety of existing bridges. The results show a good effectiveness and accuracy of the proposed methodology, opening new scenarios for the automatic defect detection on bridges, mainly aimed to support management companies' surveyors in the phase of in-situ structural inspection.