A Siamese Transformer Network for Zero-Shot Ancient Coin Classification.

Area of Science:

• Computer Vision
• Machine Learning
• Ancient Numismatics

Background:

• Ancient coin identification is a challenging problem in numismatics.
• Existing methods, primarily classification-based, struggle with rare coin issues and require retraining for new data.
• Traditional computer vision approaches are less effective than deep learning for this task.

Purpose of the Study:

• To develop a more robust method for ancient coin attribution that overcomes limitations of classification tasks.
• To leverage deep learning, specifically transformers, for improved feature extraction in coin analysis.
• To create a model that does not require exemplars of specific classes for training.

Main Methods:

• Adoption of a pairwise coin matching paradigm instead of traditional classification.
• Implementation of a Siamese neural network architecture, specifically a Double Siamese Vision Transformer (ViT).
• Utilized transfer learning on a dataset of 14,820 ancient coin images across 7605 issues, with a small training set of 542 images.

Main Results:

• The Double Siamese ViT model achieved an overall accuracy of 81%, significantly surpassing the state of the art.
• The model demonstrates effectiveness in handling issues with few or no exemplars.
• Analysis indicated that most errors were due to data quality issues rather than algorithmic limitations.

Conclusions:

• The proposed pairwise matching approach using a Double Siamese ViT is highly effective for ancient coin attribution.
• Transformer-based models with non-local attention mechanisms show promise for analyzing complex visual data like ancient coins.
• Data pre-processing and quality checking are crucial for maximizing the performance of automated numismatic analysis systems.