Multi-task deep learning for chinese porcelain classification: addressing data scarcity through synthetic image generation and transfer learning

Chinese porcelain holds immense historical and cultural value, making its accurate classification essential for archaeological research and cultural heritage preservation. Traditional classification methods rely heavily on expert analysis, which is time-consuming, subjective, and difficult to scale. This thesis investigates the application of deep learning techniques to automate and assist the multi-task identification of Song and Yuan dynasty porcelains, addressing the critical constraint of limited annotated data in archaeological contexts. Through systematic experimentation with transfer learning, synthetic data generation, and multi-task learning frameworks, we develop a comprehensive classification system targeting four essential attributes: dynasty attribution, kiln provenance, glaze classification, and vessel typology. Explainable AI analysis reveals that the multitask CNN model learns archaeologically meaningful features. These findings align with traditional identification practices, validating the model’s decision-making process. By demonstrating that machine learning can effectively encode expert knowledge while handling data constraints typical of cultural heritage domains, this thesis establishes a reproducible framework applicable to other archaeological materials, ultimately bridging computational methods with cultural heritage.