Historical building damage assessment and classification based on transfer learning

Historic buildings, one of the fundamental elements that constitute urban identity, are unique assets that reflect the cultural heritage of societies. These buildings, which represent a society's memory with their architectural values, construction techniques, and historical backgrounds, are subject to various damages due to environmental factors and human interventions. Preserving this heritage and passing it on to future generations is considered both a social and scientific responsibility. In this context, artificial intelligence-based approaches offer effective tools for documenting, analyzing, and preserving historical buildings. Artificial intelligence, particularly in image processing, has achieved significant success; deep learning methods offer powerful models capable of extracting high-level representations from data. Transfer learning, a sub-branch of these methods, enables previously trained models to be adapted to new tasks, achieving high accuracy with limited data. Model performance can be further enhanced with techniques such as Bayesian optimization. In this study, we used transfer learning-based models to automatically detect and classify damage to historical buildings. For this purpose, a unique dataset of 20,000 images was created, consisting of six classes (biological deterioration, chemical deterioration, mechanical damage, human-induced damage, material loss, and undamaged). The dataset was trained using ten different transfer learning methods (EfficientNetB3, ViT, Xception, InceptionNet, AlexNet, VGG-19, ShuffleNet, DenseNet121, MobileNet, and ResNeXt50), and Bayesian optimization was used to determine the optimal values of eight different hyperparameters. The high accuracy rates obtained demonstrate that the developed method offers an effective solution for the preservation of cultural heritage. Keywords: Historical structures, Transfer learning, Deep learning, Damage detection