TextStyleBrush: Transfer of Text Aesthetics From a Single Example.

Area of Science:

• Computer Vision
• Machine Learning
• Natural Language Processing

Background:

• Existing methods for text image analysis often require segmentation or character-level processing.
• Transferring text styles between images is challenging, especially for diverse domains like scene text and handwriting.

Purpose of the Study:

• To develop a novel self-supervised approach for disentangling text image content from appearance.
• To enable one-shot style transfer from a source text image to new content.
• To process entire word boxes without needing segmentation or character-level analysis.

Main Methods:

• A self-supervised learning framework that disentangles style and content into a fixed-dimensional vector.
• A StyleGAN-inspired approach conditioned on example style across different resolutions and content.
• Novel self-supervised training criteria using pre-trained font classifiers and text recognizers to preserve style and content.
• Introduction of the Imgur5K dataset for handwritten word images.

Main Results:

• The method successfully disentangles text style and content, enabling effective one-shot style transfer.
• Photo-realistic results are demonstrated across various text domains, including scene text and handwriting.
• Quantitative evaluations show superior performance compared to previous methods on scene text and handwriting datasets.
• A user study confirmed the effectiveness and quality of the generated text images.

Conclusions:

• The proposed method offers a robust and versatile solution for text image style and content disentanglement.
• This approach advances the state-of-the-art in text image manipulation and style transfer.
• The Imgur5K dataset provides a valuable resource for future research in handwritten text analysis.