Corrosion Assessment Using Computer Vision, Machine Learning, and Deep Learning on Imagery Data: Evaluation and Use for Bridge Inspections

Corrosion causes irreversible damage to metal structures and is one of the most common causes of bridge failure. The advent of affordable unmanned aerial vehicles (UAVs) that can carry digital cameras close to a structure presents an opportunity to quickly inspect bridges for corrosion. However, imagery-based corrosion assessment techniques must be accurate, detailed, and comprehensive to be effectively used in bridge inspections. This paper reviews and evaluates the use of imagery data and machine learning for corrosion assessment on bridges, analyzing existing research on imagery-based corrosion assessment, examining the ongoing difficulties of proposed data analysis approaches, and determining the remaining gaps in the literature. This paper evaluates the existing work with an emphasis on image processing methods; the environment and conditions in which the methods have been tested; evaluation of metrics used for performance assessment; and distinctions between corrosion detection, localization, and segmentation, as are significant in supporting bridge inspection decisions. The analysis finds that image processing techniques are typically too prone to false positives to perform well on chaotic data, particularly with nonuniform lighting and misleading objects in the image, which would be collected during bridge inspections using UAVs. Deep learning algorithms have been reported to perform more accurately than the human benchmark on similar data, and multiple researchers have developed deep learning algorithms that work well on UAV-collected data. However, there remain gaps in performance between methods tested on laboratory data compared to field data and on human-collected compared to UAV-collected data. Algorithms need to be trained on larger, more variable data sets to demonstrate generalizability across bridge inspection applications. If addressed, these methods could improve the safety, time-, and cost-efficiency of bridge inspections; help inspectors understand the state of each asset; and support structural repair, reconstruction, and prioritization decisions before failures occur.