환경별 통합에 따른 딥러닝 기반 흥인지문 지붕부 기울음 예측 실험 및 분석 = Experiments and Analyses of Inclination Prediction for the Heunginjimun Roof based on Deep Learning according to Environmental Integration

다국어 초록 (Multilingual Abstract)

Detecting the displacement of cultural heritage is essential for protecting and managing their prosperity. However, because the degree of displacement that humans can detect with the naked eye is limited and requires a lot of time and manpower, various studies are being conducted to detect the displacement and damage of cultural heritages using Deep Learning technology. In previous studies, 10 pre-trained Deep Learning neural network models through transfer learning techniques and Convolutional Neural Network (CNN) which are known for excellent for image processing were used to detect inclination in the roof of Heunginjimun. After learning to classify the images of Heunginjimun into two categories, normal and abnormal, we measured the prediction accuracy of the models generated in nine environments, classified according to season and weather. After analysis, it was raised to over 95% through hyper-parameter adjustment. In this study, among the 10 pre-trained Deep Learning Neural Network models used in previous studies, the Neural Network model with the highest accuracy was selected as the optimal model through performance verification, and then it was used to integrate 9 single environments into multiple ones. As a result, by adjusting the hyper-parameter of models below 95% in prediction accuracy, we were able to raise it above the previous percentage. Consequently, we selected the Efficientnet_v2_s model, which had the highest accuracy, to conduct an environment integration experiment. As a result, after running the learning and verifying performance in an environment that integrated all nine single environments, the roof inclination was successfully detected in all environments with an accuracy of 99.84%.