Development of Automatic Detection and Visualization System for Stone Pagoda Damage Based on Deep Learning

Abstract

Damage diagnosis through monitoring is essential for systematic and efficient conservation and management of cultural heritage. In this study, we developed a deep learning system that automatically detects and visualizes damage to stone pagodas to enable regular monitoring of cultural heritage. Mask R-CNN was used to detect and visualize damage in pixel units in stone pagoda images. A dataset specialized for stone pagodas in Korea was built and applied to train the model. The generalized performance of the trained model was evaluated on the five-story stone pagoda at Jeongnimsa Temple Site. The damage detection recall for each type was in the range of 0.86 to 0.62 based on IoU 0.50, and the damage area segmentation recall was in the range of 0.68 to 0.51. This study suggests a new safety management methodology by applying artificial intelligence to cultural heritage sites and has important applications in cultural heritage preservation.