Multi-Granularity Archaeological Dating of Chinese Bronze Dings Based on a Knowledge-Guided Relation Graph

Rixin Zhou, Jiafu Wei, Qian Zhang, Ruihua Qi, Xi Yang, Chuntao LiCVPR 2023 · DOI · arXiv

Introduction

Archaeological dating is a meticulous craft. For centuries, identifying the exact era of a Chinese bronze ding (a ritual cauldron) has required years of expert training to recognize subtle shifts in leg shapes, decorative motifs, and inscriptions. As the volume of excavated artifacts grows, the traditional "expert-only" model faces a significant bottleneck.

Evolution of Chinese bronze dings across dynasties and periods.

Researchers from Jilin University have proposed a novel solution: a deep learning framework that integrates archaeological domain knowledge through a Knowledge-Guided Relation Graph. This multi-granularity neural network automates the dating process by mapping relationships between physical attributes and historical eras.

The Challenge: Fine-Grained Visual Classification

The core difficulty lies in Fine-Grained Visual Classification (FGVC). Bronze dings from adjacent eras, such as the Late Shang and Early Western Zhou, often exhibit low inter-class variance, appearing nearly identical to the untrained eye. Conversely, dings from the same era can show high intra-class variance due to regional styles. Standard deep learning models often fail because they lack the "archaeological logic" required to prioritize specific diagnostic features, like a "phoenix pattern" or "hoof-shaped foot," over the general silhouette.

Methodology

The proposed architecture mimics the hierarchical thinking of an archaeologist through three primary modules:

  1. Multi-Granularity Module (MGM): Handles dating at both coarse (Dynasty) and fine (Period) levels. It employs a "gradient truncated addition" technique, allowing fine-grained period features to inform coarse dynasty classification without introducing noise that could degrade fine-grained performance.
  2. Knowledge Extraction Module (KEM): Acts as an "attribute spotter," trained to identify 29 distinct shapes and 96 specific characteristics, such as "animal head on the foot" or "flat title ribs."
  3. Archaeology Knowledge-Guided Relation Graph (AKG): This module builds a graph where nodes represent eras and attributes. Edges define logical relationships like subsumption (a period belongs to a dynasty) and exclusion (a ding cannot belong to two eras simultaneously). A custom Focal-type Probabilistic Loss function adaptively focuses the model on difficult decorative characteristics once basic shapes are identified.

The Bronze Ding Dataset

A significant contribution of this work is the introduction of the Bronze Ding Dataset, featuring 3,690 high-quality images spanning 4 dynasties and 11 fine-grained periods.

Example of a richly annotated bronze ding from the dataset, showing multi-granularity labels and specific characteristics.

Unlike standard datasets, this collection includes rich annotations: bounding boxes for 96 characteristics, shape labels, excavation sites, and museum locations. Mathematical analysis confirms that these attribute annotations provide significantly higher "information gain" for classification compared to popular datasets like CUB-200 or DeepFashion.

Key Results and Insights

The model achieved state-of-the-art (SOTA) performance, outperforming established methods like ConvNeXt and NTS-Net. It reached 88.79% accuracy for coarse dynasty dating and 78.83% for fine-grained period dating, surpassing the previous best multi-granularity model (HRN) by over 3% in both categories.

Qualitative analysis using Grad-CAM visualizations shows the model focusing on the exact decorative regions an archaeologist would use to justify a date. This confirms that explicitly modeling relationships between features—such as "this pattern only appeared after 1046 B.C."—is crucial for specialized fields like archaeology.

Future Directions

While the results are promising, the researchers aim to address data imbalance inherent in the archaeological record. Future work will expand the scope to other bronze vessel types (e.g., Zun, Gui) and explore multimodal integration, potentially combining visual features with chemical analysis of metal compositions for even more robust dating.

Source: Zhou et al., "Multi-Granularity Archaeological Dating of Chinese Bronze Dings Based on a Knowledge-Guided Relation Graph," arXiv:2303.15266v3, 2023.

Summary updated

BibTeX
@inproceedings{zhou2023multi,
  author = {Rixin Zhou and Jiafu Wei and Qian Zhang and Ruihua Qi and Xi Yang and Chuntao Li},
  title = {{Multi-Granularity Archaeological Dating of Chinese Bronze Dings Based on a Knowledge-Guided Relation Graph}},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2023},
  doi = {10.1109/cvpr52729.2023.00303},
}
Figure