Automatic damage detection in historical masonry structures

Examining the damages on the surfaces of historical masonry structures is laborious, costly and time consuming. In this study, automatic damage detection is proposed using deep learning based ResNet-50, VVG-16, VGG-19, Xception and Inception-V3 pretrained models to detect cracks in historical masonry structures in order to overcome these problems. In addition, class activation map approach (Grad-CAM, Grad-CAM++, Score-CAM) was used to make deep learning techniques more explainable on image-based data in order to detect the location of cracks in historical masonry structures. In the training of pre-trained ready-made models, 502 cracked and 502 crack-free images obtained from the Santa Ruins, which is located in the Dumanlı Village of Yağmurdere Sub-district, approximately 82 kilometers away from Gümüşhane city center, and which is under protection as a historical structure, were used. Among the proposed deep learning architectures, the ResNet-50 model and all visualization strategies were found to have the best performance in terms of localizing crack regions. As a result, the proposed method is a fast, efficient and reliable method for damage detection of historical masonry structures and it is possible to apply these methods automatically. This will help preserve historic masonry structures and make them usable in the future.