Autonomous Bridge detection based on ResNet for multiple damage types

Guojin Tan,Zheng Yang
DOI: https://doi.org/10.1109/cyber53097.2021.9588187
2021-01-01
Abstract:Early and timely detection of bridge surface damage plays a key role to maintaining bridge functionality and guaranteeing transportation safety. The common practice is to detect bridge damage by artificial visual inspection. However, it is time-consuming, labor-intensive and has a low-efficiency. This paper presents a simple, high-accuracy framework to detect four types of bridge damage based on ResNet-101 model. A damage images dataset with 2,726 images including crack, spall, rebar exposed and honeycomb is developed for training and testing the ResNet-101 model. To enhance detection accuracy, a transfer learning method with fully pre-trained weights from a similar dataset was adopted as it improves the original ResNet-101. The robustness and adaptability of the trained ResNet-101 model are tested on test dataset. Result shows that compared with traditional classification models, the proposed approach can effectively detect those four kinds of bridge damage in realistic situations with a better performance.
What problem does this paper attempt to address?