Crack Segmentation on Earthen Heritage Site Surfaces

Abstract

Earthen heritage sites are historical relics left by ancient human activity, with earthen as the primary building material, and have significant historical, scientific, and artistic value. However, many sites have experienced extensive deterioration caused by environmental forces and human factors. A crack is a kind of typical damage to the walls of earthen heritage sites. Studies of the crack-formation process can effectively predict trends in damage, which will play a critical role in the maintenance of earthen heritage sites. This study is the first of its kind to propose a deep learning method to study the cracks on earthen heritage sites at the pixel-level, adopt the idea of transfer learning, and employ a mixed-crack image dataset for training three deep learning models. The precision, recall, IoU, and F1 metrics were used to evaluate the performance of the trained models. The experimental results showed that FPN-vgg16 appeared to have the highest level of applicability to detect cracks on earthen heritage sites among all networks, due to the highest F1 score of 84.40% and the highest IoU score of 73.11%. The results illustrated that the proposed method in this paper can effectively be used to analyze the rammed earth surface crack images, with great potential in related research fields.