What question did this study set out to answer?

This research aims to automate the classification of Chinese porcelain artifacts using deep learning techniques.

February 8, 2026

Multi-task learning for identification of porcelain in Song and Yuan dynasties

Key Points

This research aims to automate the classification of Chinese porcelain artifacts using deep learning techniques.
Applied multi-task learning and transfer learning methods for classification tasks
Evaluated performance of four Convolutional Neural Networks: ResNet50, MobileNetV2, VGG16, InceptionV3
Compared models with and without pre-trained weights
Implemented eXplainable AI techniques for interpretability of results
MobileNetV2 achieved 93.5% accuracy, outperforming other models
ResNet50 and InceptionV3 followed with accuracies of 92.875% and 92.725%, respectively
VGG16 underperformed with 84.3% accuracy
Highest overall accuracy in dynasty classification at 96.775%
Transfer learning provided gains in accuracy up to 12% across tasks

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 Deep Learning (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. MobileNetV2 achieves the best average performance (93.5% accuracy), followed closely by ResNet50 (92.875%) and InceptionV3 (92.725%), while VGG16 underperforms (84.3%). Generally, the models achieve their highest accuracy in dynasty classification (96.775% on average), whereas glaze and ware are slightly different tasks (92.8% and 92.65%), and type classification proves most challenging (81.175% on average). Results demonstrate that transfer learning consistently improves classification accuracy, with overall gains until 12% across tasks. Furthermore, we incorporate eXplainable AI (XAI) techniques, including Grad-CAM and a SHAP-inspired framework, to reveal the visual cues driving model decisions and to ensure interpretability in cultural heritage applications.

Mark Helpful

Bookmark

Relay