One shot ancient character recognition with siamese similarity network.

Area of Science:

• Digital Humanities
• Computer Vision
• Artificial Intelligence

Background:

• Ancient character recognition is crucial for historical studies and cultural heritage.
• Challenges include limited data, class imbalance, glyph diversity, and open-set recognition.

Purpose of the Study:

• To develop an efficient ancient character recognition system using similarity learning.
• To overcome common challenges in ancient character datasets and recognition tasks.

Main Methods:

• Proposed a Siamese similarity network with a multi-scale fusion backbone and embedded structure for feature extraction.
• Introduced a novel soft similarity contrast loss function to prevent overfitting and improve optimization.
• Developed a cumulative class prototype for robust class representation.

Main Results:

• The model demonstrated high-efficiency discriminative performance in ancient character recognition.
• Achieved superior results compared to traditional deep learning and classic one-shot learning methods.
• Successfully handled the rejection of unknown categories, enabling recognition of newly discovered characters.

Conclusions:

• The proposed Siamese similarity network effectively addresses challenges in ancient character recognition.
• The novel loss function and prototype method enhance model robustness and accuracy.
• This approach offers a significant advancement for the study of ancient scripts and cultural heritage preservation.