A study on utilization of deep neural networks for corrosion detection in steel girder bridges in Fukushima Prefecture

Abstract

In recent years, the number of aging bridges has increased. Since the number of bridge inspection engineers is to decrease, future inspections require labor-saving and streamlined operations. This study developed a learning model for corrosion detection in steel girder bridges using training data based on bridge inspection results from Fukushima Prefecture in 2016. The model was employed to four deep neural network architectures and their generalization performance was compared. As a result, corrosion and weathering of steel and concrete classes achieved the highest classification accuracy using AlexNet, MobileNetV2, and ResNet18 within the scope of this study. Furthermore, the number of training data was positively correlated with the classification accuracy of both corrosion and weathering of steel classes in almost all learning models. Therefore, it was observed that these classes significantly contributed to classification accuracy with an increase in the number of training data.