A Study on Solving the Data Imbalance Problem for Detecting Heunginjimun Roof Tilt Using Transfer Learning Algorithms

Abstract

Cultural heritage with high historical value requires continuous management and protection. However, recognizing subtle changes with the naked eye has limitations and requires much time and personnel deployment. To solve this problem, we will automatically detect the tilt of Heunginjimun's roof using Transfer Learning algorithms. In a previous study, among single environments classified into nine types according to season and weather, the ratio of normal and abnormal images in the winter/night and winter/daytime datasets was unbalanced at 9:1 and 8:2. As a result, problems with poor prediction accuracy occurred in some experiments. In this paper, to solve this problem, we adjusted the composition ratio of the dataset and measured the prediction accuracy. When comparing the measurement results with previous studies, the dataset size was reduced by half, but the accuracy was higher. This showed that higher accuracy and performance can be expected by achieving the balance between classes rather than increasing the dataset size.