Automatic recognition of damages in cultural heritage building images using an encoder-decoder network

The widespread presence of damaged cultural heritage buildings highlights the urgent need for timely detection and the implementation of appropriate protective measures. However, manually analyzing a large volume of damage in cultural heritage buildings is not only time-consuming but also heavily dependent on the expertise of the practitioner. To address this challenge, this paper proposes an automatic recognition algorithm based on an encoder-decoder network (DeepLabV3+) to segment various types of damage in cultural heritage buildings, including plant, cracks/discontinuous, spalling, efflorescence, erosion, etc. To enhance the recognition capability of the proposed algorithm, a Grid-Mask approach is applied for data augmentation on the established cultural heritage damage dataset. Preliminary results from two field experiments demonstrate that the proposed recognition algorithm can accurately identify and classify damage in cultural heritage buildings.