Enhanced Generative Structure Prior for Chinese Text Image Super-Resolution.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Digital Image Processing

Background:

• Super-resolution (SR) for text images is difficult due to character variations.
• Existing SR methods often neglect complex scripts like Chinese.
• Restoring precise character strokes in low-resolution (LR) text is a significant challenge.

Purpose of the Study:

• To develop a high-quality text image SR framework for Chinese characters.
• To introduce a novel structure prior for enhanced visual quality in text SR.
• To address limitations of existing methods focusing on English text.

Main Methods:

• Proposed a novel structure prior for structure-level guidance in text SR.
• Integrated the structure prior into a StyleGAN model for restoration.
• Implemented a codebook-based mechanism to control character structure and style variations.
• Leveraged collaborative interaction between codebook and style vectors ($w$) for high-resolution prior generation.

Main Results:

• The structure prior provides robust, character-specific guidance for SR.
• Accurate restoration of clear strokes in degraded Chinese characters was achieved.
• The framework effectively handles real-world LR Chinese text with irregular layouts.
• Improved visual quality and stroke integrity in super-resolved text images.

Conclusions:

• The proposed framework successfully restores high-quality Chinese text images from LR inputs.
• The novel structure prior significantly enhances the performance of text SR.
• The method is effective for complex scripts and irregular text layouts, advancing the field of text image SR.