Dual-backbone fusion network for damage segmentation in cultural heritage buildings

Cultural heritage (CH) buildings are vulnerable to damage due to aging and environmental factors, necessitating timely detection and maintenance. This paper proposes a lightweight dual-backbone segmentation model for damage detection in CH structures. The architecture integrates a Swin Transformer branch to capture global contextual information and a YOLOv8-Ghost branch to preserve fine-grained local details, with a Content-Guided Attention (CGA) fusion mechanism employed to enhance inter-channel feature interactions. A five-class Roman amphitheater damage dataset with 2010 images was constructed for training and evaluation. The proposed model is applied to damage detection in the Arena, Verona, Italy, which experienced local collapse accident on January 23, 2023. Experimental results demonstrate that the model achieves robust segmentation performance under challenging conditions such as low lighting, occlusions, and heterogeneous surface textures. The inspection results of both the exterior and interior facades of the Arena confirm the effectiveness and efficiency of the proposed dual-backbone fusion strategy.