Automated bridge crack evaluation through deep super resolution network-based hybrid image matching

Abstract

This article proposes a deep super resolution crack network (SrcNet)-based automated bridge crack evaluation technique through hybrid image matching. The hybrid images combining vision and laser-induced infrared (IR) thermography images are able to improve crack detectability while minimizing false alarms. However, matching of hybrid image is difficult to equivalently evaluate cracks due to their different image resolutions. To resolve the technical difficulty, the SrcNet-based hybrid image matching technique is newly proposed and experimentally validated through in-situ bridge tests. First, a hybrid crack evaluation system is developed. Then, SrcNet is established to make the vision and IR image resolutions equivalent. Using the resolution-modified vision and IR images 100 μm-level cracks are sophisticatedly evaluated in a pixel-level through hybrid image matching. The validation test results reveal that false alarms caused by harsh bridge condition are effectively reduced up to 49.91 and 13.31% in terms of precision and recall, respectively.