CFOR: Character-First Open-Set Text Recognition via Context-Free Learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Open-set text recognition models struggle with novel characters, indicating biases in training data.
• Existing models often fail to generalize to unseen characters and scripts.

Purpose of the Study:

• To develop a robust open-set text recognition framework that can identify novel characters.
• To mitigate biases in language models by learning context-free character representations.

Main Methods:

• Proposed a Character-First Open-Set Text Recognition framework with two cotrained, context-free learning tasks.
• Implemented Context Isolation Learning using a weakly supervised character mask to remove contextual information.
• Introduced Individual Character Learning with synthetic samples for single-character classification.

Main Results:

• The framework successfully recognized unseen characters in Japanese, Korean, and Greek without retraining.
• Achieved over 64% F1-score for spotting unseen Japanese characters.
• Demonstrated 91.5% line accuracy on the IIIT5k dataset with a speed of over 69 FPS.

Conclusions:

• The Character-First framework offers a universal and lightweight OCR solution for both open-set and close-set scenarios.
• Learning context-free character representations effectively addresses the challenge of novel character recognition.
• The model exhibits strong generalization capabilities across diverse scripts and character sets.