A dataset of oracle characters for benchmarking machine learning algorithms

Area of Science:

• Digital Humanities
• Computer Vision
• Artificial Intelligence

Background:

• Oracle bone script offers insights into ancient Chinese culture, history, and language.
• Existing datasets like MNIST lack the complexity of real-world ancient script challenges.

Purpose of the Study:

• Introduce the Oracle-MNIST dataset for benchmarking pattern classification tasks.
• Provide a more challenging alternative to MNIST for machine learning research.
• Facilitate research on classifying ancient scripts with inherent noise and style variations.

Main Methods:

• Curated a dataset of 30,222 grayscale images (28x28) of ancient Chinese characters from 10 categories.
• Structured the dataset (27,222 training, 300 test per class) for compatibility with existing systems.
• Focused on realistic challenges including significant noise and diverse writing styles.

Main Results:

• The Oracle-MNIST dataset presents a more difficult classification task than the original MNIST.
• The dataset captures the unique noise and stylistic variations inherent in ancient scripts.
• Enables direct comparison and integration with existing machine learning classifiers.

Conclusions:

• Oracle-MNIST serves as a valuable benchmark for pattern classification of historical artifacts.
• The dataset advances machine learning research by incorporating realistic challenges of ancient data.
• Promotes further investigation into the interpretation of ancient writing systems through AI.