DenseNet with Attention Mechanism for Ancestral Hall Architectural Style Classification and Decorative Element Recognition in Pearl River Delta
Year: 2026
Type: conference-paper
Abstract: Leveraging publicly available architectural images of approximately 2000 ancestral halls (covering the Guangfu, Hakka, and Teochew communities) from the Guangdong Provincial Cultural Heritage Bureau and various municipal databases, this paper proposes an improved DenseNet-121 model for their analysis. The model incorporates a Convolutional Block Attention Module (CBAM) for channel and spatial feature refinement and a multi-label classification head. Its key innovation lies in a hierarchical classification strategy: first identifying the regional architectural style and then recognizing decorative elements such as wood carvings, lime sculptures, and murals. The model achieved a style classification accuracy of 92.6% and a mean Average Precision ([email protected]) of 86.3% for decorative element recognition. This approach demonstrates significant potential for application in the digital archiving of historical buildings and intelligent cultural tourism guidance systems.
Authors: Yunlei Chen, Bo Wang, Zhen Zhang, Jialang Zhang
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7208740236