A Study of the Use of Deep Learning to Detect Corrosion and Damage in Road facilities

Application of CNN and MT Method

Abstract

In the maintenance of equipment, inspection and replacement receive major attention because of their direct bearing on accidents. When there are many things to be inspected, however, the cost and labor involved become burdensome, so there is a desire to replace sole reliance on visual inspection with more efficient methods, such as methods applying imaging and AI. Here we use CNN and Residual network （ResNet） to compress the information in images captured during inspection, and apply quality engineering methods to the compressed data to express departure from normal in terms of Mahalanobis distance. The result is that although slight errors of judgment are observed, the usefulness of CNN and ResNet image processing becomes clear, and the extent of damage can be determined by evaluation in terms of Mahalanobis distance based on CNN, suggesting that the Mahalanobis distance can become a useful tool in equipment maintenance and accident prevention.