Displacement Detection of Wooden Cultural Properties Using Unsupervised Learning
Jungwoo Park, Sang‐Yun Lee · 2022 IEEE International Conference on Consumer Electronics-Asia (ICCE-Asia) · 2022
Since supervised learning requires expert labeling work that requires a lot of time and money to acquire data, an alternative is required in the field of cultural property management. In addition, the existing method based on contact type sensors to detect displacement occurring in wooden cultural heritage has a risk of damage to cultural heritage. This paper proposes to apply an artificial intelligence model using f-AnoGAN, which shows good performance in detecting abnormalies, to overcome the difficulty of obtaining abnormal data in wooden cultural assets with an unsupervised learning approach and to explore alternatives to the conventional contact type sensor-based method. The applied anomaly detection model, f-AnoGAN, is characterized by the fact that learning is performed to simultaneously minimize the error in the image space and the error in the feature space when a new image is mapped to the latent space and then returned to the image space. Our experimental results show that the f-AnoGAN model successfully performs judgment on whether or not displacement has occurred in wooden cultural properties using actual CCTV color images, and further presents the location of displacement when displacement occurs.