Deep Learning Approaches for Image-Based Detection and Classification of Structural Defects in Bridges

Abstract

The paper presents a study about the defect detection on structural elements of existing reinforced concrete bridges through a machine-learning approach. In detail, the proposed methodology aims to explore the possibility of automatically recognising deficiencies on bridges’ elements, e.g., cracks, humidity, by employing a training of existing convolutional neural networks on a set of photos. The initial database, characterized by 2.436 images, has been firstly selected and after has been 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.