Evaluation and Prediction of Steel Corrosion Using Visible Light Images and Machine Learning

Abstract

To maintain the safety of social infrastructures, it is crucial to evaluate the corrosion state of materials and predict their lifespan. We aimed to make it easier by applying machine learning to visible light images of materials. In the present report, machine learning was applied to images of carbon steel specimens prepared by salt spray, to classify the corrosion products on the specimen surface and to predict their change over time. It was shown that the images could be reconstructed into label maps corresponding to the corrosion products on the surface by unsupervised learning on the pixel colour values, and that changes of the label distributions over time could also be predicted almost accurately by supervised learning. The results demonstrated the possibility of an easy and accurate method to evaluate and predict corrosion state of infrastructures using the combination of machine learning and visible light images taken by ordinary cameras.