GAN based Deep Learning Model for Detecting Damage and Displacement of Cultural Asset
Woo-Yeon Choi, Jeong‐Woo Park, Sang‐Yun Lee · 2021
The following paper proposes F-AnoGan model for anomaly detection regardless of the types of damage or displacement in wooden cultural assets image. Firstly, by training WGAN’s Generator model and Discriminator model using Normal images, we can get the value of parameter that is good at creating fake normal images. After the WGAN’s parameter is fixed, the Encoder is trained by the normal image that was used to train. Then, we input the new image data of wooden cultural assets to well-trained WGAN and Encoder and calculate Anomaly Score to discriminate and detect area where anomaly exists. The virtual dataset is established with normal data of cultural assets collected and abnormal data that contains anomaly made using Photoshop. The final experimental results confirmed that model cannot detect minute abnormal regions of abnormal images but accurately determines whether it is normal or abnormal. The proposed method from these results show that it is suitable for future anomaly detection of damage and displacement in wooden cultural assets.