Enhanced deep learning for detecting Chinese vernacular architectural roofs with complex forms

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

The roof is one of the most iconic elements of Chinese vernacular architecture. Accurate and efficient recognition of roof types is essential for the digital documentation, dynamic monitoring, and restoration planning for vernacular architectural heritage. This study proposes a novel object detection technique by integrating the DCNv2, the DyHead and the SIoU to address challenges such as complex morphology, diverse materials, and feature occlusion in roof images. Experiments indicate that YOLOv10-DDS achieves 80.8% and 62.4% on mAP50 and mAP50-95, improving by 2.1% and 1.5% over the baseline model and outperforming other advanced methods. Validation across diverse scenarios confirms that the model maintains high accuracy and adaptability. A three-tier classification system comprising 16 subclasses of Chinese vernacular architectural roofs and a dual-perspective dataset were developed. This study establishes a technical foundation for the preservation and restoration of vernacular architectural heritage, expanding the application scope of object detection in cultural heritage research.