Few-shot bronze vessel classification via siamese fourier networks

Area of Science:

• Archaeology
• Computer Science
• Materials Science

Background:

• Ancient Chinese bronze vessels are vital for understanding East Asian technological development.
• Identifying intricate and periodic patterns on these artifacts is challenging for human vision and conventional RGB methods.
• Existing methods struggle to capture the regular textures characteristic of bronze vessel designs.

Purpose of the Study:

• To propose a novel deep learning model for few-shot regular pattern classification on ancient Chinese bronze vessels.
• To address the limitations of RGB-domain methods in capturing periodic designs.
• To enhance the analysis of East Asian technological history through improved artifact pattern identification.

Main Methods:

• Development of Siamese Fourier Networks (SFN), a parallel network model.
• Integration of Siamese networks for shape differentiation and Fourier features for texture extraction.
• Optimization of parallel networks using a combination of BCE loss and focal contrastive loss for sample balancing.
• Creation of the Bronze Vessel Dataset with 527 diverse and unbalanced samples.

Main Results:

• Demonstrated the superiority of SFN over advanced few-shot methods.
• Achieved significant improvements in accuracy for regular pattern classification.
• Validated the effectiveness of the focal mechanism in optimizing the network.

Conclusions:

• SFN offers a powerful solution for few-shot regular pattern classification, particularly for complex artifact analysis.
• The proposed method enhances the study of ancient Chinese bronze vessels and East Asian technological advancements.
• Fourier features and focal loss are effective in handling intricate designs and imbalanced datasets.