VISUALIZATION AND UNDERSTANDING OF CONVOLUTIONAL NEURAL NETWORK FOR DAMAGE LEVEL DETERMINATION OF BRIDGE MEMBERS

Abstract

In the periodic bridge inspection in Japan, engineers are required to determine the damage level for each structural member by visual inspection in all of seven hundred thousand bridges. More sustainable and low cost bridge inspection system is required in the future. For this purpose in this study, the applicability of deep machene learning to the damage level determination of bridge members is verified based on the visualization technique. In detail, the convolutional neural network (CNN) that determines the damage level of the bridge member from the image data was constructed. And then, the Gradient-weighted Class Activation Mapping (Grad-CAM) was applied to verify the features of image, which contribute on the damage level determination. By comparing the heat-maps of the Grad-CAM, the consistency of the feature to determine the damage level in the CNN to the feature used in the inspection conducted by expert engineers could be discussed. Furthermore, it was also shown that knowledge from the outputs of Grad-CAM is applicable to the improvement and encourage of acceptance of the CNN.