Architectural Heritage Recognition via Computer Vision: Case Study on Traditional Chinese Window Pattern Detection in Suzhou Gardens

Key words: Computer Vision, Suzhou Garden, Traditional Chinese Window, Architectural Heritage, Object Detection

Introduction
In traditional Suzhou gardens, windows—particularly their lattices—hold significant aesthetic and cultural value. These intricate and diverse designs are challenging to identify quickly without specialist knowledge. Visual AI models such as You Only Look Once (YOLO) [1] and Graph Neural Networks (GNNs) [2] have been increasingly applied to historic architecture for efficient recognition of complex patterns, including detecting traditional Chinese roofs [3] and classifying building types in Athens [4]. Other AI approaches have supported heritage studies on religious buildings and monuments . YOLO is notable for real-time, accurate detection, suitable for mobile and desktop use. However, automated detection of traditional Chinese window lattices remains underexplored, with most studies focusing only on locating windows rather than classifying lattice patterns. This study presents Heritage-Scan (H-Scan), a YOLOv8n-based software for detecting long-window lattices in Ming (1368–1644) and Qing (1636–1912) dynasty gardens, addressing:

1. How effective is a YOLOv8n-based object detection model in accurately identifying diverse traditional long-window lattice patterns in Chinese gardens?
2. How can automated pattern recognition enhance access to and understanding of traditional Chinese architectural elements in digital heritage education?

Methodology
A dataset of 1,143 images was compiled, with 315 classified into 14 lattice types following Yuanyan (1631). YOLOv8n was trained for 50 epochs (640 image size, batch 8), achieving baseline mAP50 = 0.59 and precision = 0.49. Fine-tuning (lr0 = 0.0005) improved mAP50 to ~0.66 and precision to ~0.77. Class imbalance remained a challenge. The model was integrated into H-Scan, supporting real-time and batch detection, automatic annotation, and Excel export.
Results and Conclusions
The fine-tuned model achieved strong performance (average 0.30 ms per frame), with high accuracy for some categories (Haitang lattice mAP50 = 0.803) but misclassifications in low-salience patterns. Culturally, the study provides a heritage dataset of over 1,000 images; technologically, it delivers an efficient detection model within a functional application, enabling practical AI-assisted recognition for preservation, education, and public engagement.


References

[1] Redmon J, Divvala S, Girshick R, Farhadi A. You only look once: Unified, real-time object detection. In: Proc IEEE Conf Comput Vis Pattern Recognit (CVPR). 2016; 779–88.

[2] Jiang B, Chen S, Wang B, Luo B. MGLNN: semi-supervised learning via multiple graph cooperative learning neural networks. Neur Netw. 2022; 153:204–14.

[3] Hou M, Hao W, Dong Y, Ji Y. A detection method for the ridge beast based on improved YOLOv3 algorithm. Herit Sci. 2023;11(1):167.

[4] Janković R. Machine learning models for cultural heritage image classification: comparison based on attribute selection. Information. 2019;11(1):12.