Multi-task Learning for Identification of Porcelain in Song and Yuan Dynasties

TL;DR

This work tackles automated classification of Song- and Yuan-dynasty porcelain across dynasty, glaze, ware, and type using a multi-task CNN framework with transfer learning. By evaluating ResNet50, MobileNetV2, VGG16, and InceptionV3, the study shows that pre-trained weights substantially boost performance, with MobileNetV2 and ResNet50 offering the best balance across tasks, though the Type classification remains the most challenging. The dataset comprises 5,993 images from the National Palace Museum Open Data, exhibiting class imbalance and multiple viewing angles, which informs the analysis of balanced accuracy and model robustness. The findings highlight the value of transfer learning in cultural heritage applications and point to future directions such as domain-specific pre-training, attention mechanisms, explainable AI, and generalization to additional artifact classes.

Abstract

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 paper explores the application of DL and transfer learning techniques to automate the classification of porcelain artifacts across four key attributes: dynasty, glaze, ware, and type. We evaluate four Convolutional Neural Networks (CNNs) - ResNet50, MobileNetV2, VGG16, and InceptionV3 - comparing their performance with and without pre-trained weights. Our results demonstrate that transfer learning significantly enhances classification accuracy, particularly for complex tasks like type classification, where models trained from scratch exhibit lower performance. MobileNetV2 and ResNet50 consistently achieve high accuracy and robustness across all tasks, while VGG16 struggles with more diverse classifications. We further discuss the impact of dataset limitations and propose future directions, including domain-specific pre-training, integration of attention mechanisms, explainable AI methods, and generalization to other cultural artifacts.

Figures (10)

• Figure 1: Images of the same white porcelain pillow (Yuan dynasty, Peng ware) from multiple angles.
• Figure 2: Network Structure of the Multi-Task Model
• Figure 3: Training Loss Comparison
• Figure 4: Accuracy of Porcelain Classification Across Four Models
• Figure 5: Confusion Matrix of Four Tasks in InceptionV3