FAIR-Net: A Fuzzy Autoencoder and Interpretable Rule-Based Network for Ancient Chinese Character Recognition.

Area of Science:

• Computer Science
• Artificial Intelligence
• Digital Humanities

Background:

• Ancient Chinese scripts face degradation, hindering manual transcription and conventional OCR.
• Existing OCR methods struggle with historical text due to erosion, damage, and stylistic variations.

Purpose of the Study:

• To develop an interpretable and efficient AI model for recognizing degraded ancient Chinese characters.
• To improve the digitization and preservation of historical Chinese documents.

Main Methods:

• Proposed FAIR-Net, a hybrid architecture merging unsupervised deep autoencoders for feature learning with fuzzy rule-based classification.
• Utilized Fuzzy C-Means (FCM) for soft clustering and Iteratively Reweighted Least Squares Estimation (IRLSE) with Softmax for transparent predictions.
• Constrained model weights as linear mappings to ensure interpretability.

Main Results:

• Achieved 97.91% accuracy on benchmark datasets, significantly outperforming baseline CNNs with high statistical significance.
• Demonstrated improved processing efficiency, reducing inference time by up to 98.9% compared to other models.
• Showcased robustness on a large ancient Chinese character dataset (83.25% accuracy) and confirmed enhanced resilience to glyph ambiguities via fuzzy rule visualization.

Conclusions:

• FAIR-Net offers a practical, interpretable, and efficient solution for ancient Chinese character recognition.
• The model's transparency and performance facilitate the digitization and preservation of invaluable historical corpora.
• FAIR-Net's hybrid approach advances AI applications in historical linguistics and cultural heritage preservation.