Can Deep Learning Identify Early Chinese Ceramics Using Only 2D Images?

Area of Science:

• Archaeology
• Art History
• Computer Science

Background:

• Early Chinese ceramics are vital for understanding historical developments.
• Deep learning offers new possibilities for analyzing historical artifacts like ceramics.
• Identifying and dating ceramics often requires specialized domain knowledge.

Purpose of the Study:

• To investigate the efficacy of deep learning in identifying early Chinese ceramics using 2D images.
• To assess the performance of state-of-the-art neural networks on ceramic visual feature recognition and dating.
• To develop a class-imbalance learning strategy for multi-label ceramic identification tasks.

Main Methods:

• Collected a diverse dataset of ancient Chinese ceramics spanning 15 dynasties.
• Utilized five state-of-the-art neural networks for analysis.
• Implemented a class-imbalance learning strategy to address dataset biases.
• Evaluated models on ceramic visual feature recognition and dating tasks.

Main Results:

• Deep learning models achieved high accuracy in recognizing visual features like glaze and shape.
• Ceramic dating was feasible for major dynasties but challenging overall.
• Cultural and artistic continuity can lead to misclassifications in dating.
• AI demonstrated potential in identifying unlabeled, time-relevant features.

Conclusions:

• Deep learning is effective for visual feature recognition of early Chinese ceramics.
• AI shows promise for dating ceramics, though challenges remain.
• AI can aid in understanding the evolution and inheritance of ceramic styles.
• Further research can leverage AI to uncover hidden patterns in historical artifacts.