Learning-Based Ordering Characters on Ancient Document.

Area of Science:

• Computer Science
• Artificial Intelligence
• Document Image Analysis

Background:

• Character recognition in scanned documents is hindered by detection order inaccuracies.
• Organizing detected characters is crucial for accurate translation, especially for documents with variations like antique handwritten texts.

Purpose of the Study:

• To propose a learning-based model for ordering characters from documents with diverse variations, improving translation accuracy.
• To address the challenge of character ordering in documents with non-uniform layouts, such as antique handwritten materials.

Main Methods:

• A learning-based model was developed to learn character ordering operations from data.
• The model utilizes a network architecture with an expanded receptive field to handle long-range dependencies.
• A ground truth assignment method based on absolute position and Region of Interest (ROI) was employed to improve accuracy.

Main Results:

• The proposed method achieved a modified edit distance 0.43 times that of human-designed algorithms.
• The modified Fisher criterion demonstrated a 1.46 times improvement compared to human-designed algorithms.
• The model effectively handles variations in character number, area, and position.

Conclusions:

• The learning-based approach significantly outperforms human-designed algorithms in character ordering for diverse document images.
• The method provides a robust solution for digitalizing and translating complex historical and varied documents.
• Accurate character ordering is essential for reliable document digitalization and translation.