Multi-view dynastic classification and visual interpretation of blue-and-white porcelain

Key Points

• The central aim is to classify and interpret blue-and-white porcelain based on dynastic periods using deep learning.
• Curated a multi-view dataset of Ming and Qing porcelain with 284 objects and 963 images.
• Employed a pretrained and fine-tuned ResNet-50 model for classification.
• Incorporated multiple object views during training to enhance model performance.
• Utilized visual explanation techniques like Grad-CAM and t-SNE for feature interpretation.
• The ResNet-50 model achieved the highest classification performance among evaluated models.
• Incorporating multiple views significantly improved classification accuracy compared to single view training.
• Visual explanation techniques revealed shape and motifs as key characteristics.
• t-SNE visualizations indicated global clustering by dynasty and local clustering by object.

Abstract