CONSIDERATION OF DEFECT EXTRACTION USING AI FOR SUBWAY TUNNELS

Abstract

 Tokyo Metro conducted a study to improve the efficiency of hammering tests to prevent concrete spalling in subway tunnels. The study focused on evaluating the accuracy of AI-based defect detection using different types of tunnel images. Two types of images—high-saturation color images and grayscale images—were used to create defect distribution maps, which served as ground truth data for training AI models. Accuracy tests were performed using various combinations, including applying a model trained on high-saturation color images to analyze grayscale images. The results showed that detection accuracy tended to be higher when the image types used for training and evaluation were the same. Furthermore, the study confirmed demonstrated that even grayscale images, which lack color information and generally have lower resolution, can still be used effectively for defect detection with sufficient training. These findings indicate that further research on grayscale image-based AI detection is worthwhile.