Semi-Supervised Scene Text Recognition.

Area of Science:

• Computer Vision
• Machine Learning
• Natural Language Processing

Background:

• Supervised methods dominate scene text recognition but require costly labeled data.
• Character-level or pixel-wise annotations are often needed, increasing data preparation effort.
• Unlabeled data is abundant, especially for low-resource languages, presenting an opportunity for improvement.

Purpose of the Study:

• To develop a novel semi-supervised method for scene text recognition.
• To reduce reliance on expensive labeled data by leveraging unlabeled data.
• To improve recognition performance and reduce annotation effort.

Main Methods:

• Proposed a semi-supervised approach for scene text recognition.
• Introduced two global metrics: edit reward and embedding reward.
• Utilized reinforcement learning to optimize these rewards for generated text quality.

Main Results:

• The edit reward measures the discrepancy between ground truth and generated text.
• The embedding reward assesses similarity between image and text features in a common space.
• Experimental evaluations on challenging benchmarks demonstrated significant effectiveness.

Conclusions:

• The proposed semi-supervised method effectively utilizes unlabeled data for scene text recognition.
• The approach significantly reduces annotation effort while maintaining competitive performance.
• This method offers a viable solution for low-resource languages and reduces data labeling costs.