Damage Detection and Safety Diagnosis for Immovable Cultural Assets Using Deep Learning Framework

Abstract

Cultural assets are recognized as having cultural values among the products of the cultural activities our ancestors carried out and of great historical or artistic value as well. Since cultural assets, which are also those of the nation, cannot be restored once damaged, both preservation and transmission of the cultural assets are critical. When some changes occur in cultural properties by any chance, if they can be identified as quickly as possible by using Deep Learning, rapid initial response and management will be possible, which will significantly help the management and preservation of cultural properties. This paper aims to create a Deep LearningFramework that can effectively detect minute inclinations that may be difficult to identify with human eyes. For Heunginjimun, the No.1 national treasure of South Korea, four types of pretraining-based Deep Learning models were used, which were EfficientNetB0, EfficientNetB2, ShuffleNet_v2, and AlexNet. The experiment was conducted using a typical dataset in the environment constructed by using Heunginjimun CCTV images, and then an abnormal dataset was regenerated based on the typical dataset. As a result of applying the Deep Learning model to each environment we have built, the average prediction accuracies of EfficientNetB0, EfficientNetB2, and ShuffleNet_v2 among the four models were 99.69%, 99.66%, and 93.46%, respectively, showing high prediction accuracies. Therefore, it is judged that the results of the experiment will be of great help in the field of management and preservation of cultural properties in the future.