Science & Technology

Facebook AI can copy photo text styles from one word

Improve your enterprise data technology and strategy with 2021 transformation..

Facebook introduced today TextStyleBrush, An AI research project that allows you to copy the text style of a photo from one word. The company claims that TextStyleBrush, which allows you to edit and replace any text in an image, is the first “unsupervised” system of its kind to recognize both typefaces and handwriting.

AI-generated images are evolving rapidly and have obvious business applications such as photorealistic translations of languages. Augmented Reality (AR).. (The AR market was expected to be worth $ 18.8 billion by the end of 2020, according to However, building a system that is flexible enough to understand the nuances of text and handwriting is not only about typography and calligraphy, but also about transforming rotating, curved text, deformations, background clutter, images, etc. It’s a difficult task because it means understanding the style of. noise.

According to Facebook, TextStyleBrush works the same way the Style Brush tool works in a word processor, but for the aesthetics of the text in an image. Unlike previous approaches that define specific parameters such as typeface and target style monitoring, it takes a more comprehensive training approach to unravel the content of the text image from all aspects of its appearance.

Unsupervised learning

The “unsupervised” part of the system refers to unsupervised learning. This is the process by which your system is exposed to “unknown” data that does not have previously defined categories or labels. TextStyleBrush had to teach itself to classify data, process unlabeled data and learn from its unique structure.

As Facebook points out, systems like TextStyleBrusht usually involve training with annotated data that teaches the system to classify individual pixels as “foreground” or “background” objects. However, it is difficult to apply this to images captured in the real world. The width of the handwriting can be less than 1 pixel and the foreground and background should be labeled to collect high quality training data.

In contrast, given a detected “textbox” containing the source style, TextStyleBrush uses a single sample to render the new content in the style of the source text. While you may struggle with metallic objects and text written in different colors, Facebook says TextStyleBrush proves that you can build a system that allows you to transfer the aesthetics of your text more flexibly than before.

Facebook TextStyle Brush

“I hope this work will continue to lower the barriers to photorealistic translation. [and] “Creative self-expression,” Facebook said in a blog post. “This technology is research, but it translates text in images into different languages, creates personalized messages and captions, and someday uses AR to facilitate the actual translation of road signs. , In the future we can enhance various useful applications. “

Here are the features, methods, and results of working with TextStyleBrush: Available On Facebook’s developer portal. According to the company, it will be posted in peer-reviewed journals in the future.

Venture Beat

VentureBeat’s mission is to become a digital town square for technical decision makers to acquire knowledge about innovative technology and trading. Our site provides important information about data technologies and strategies to guide you when you lead your organization. We encourage you to become a member of the community and access:

  • The latest information on the subject you are interested in
  • Newsletter
  • Gated sort reader content and discounted access to valuable events such as: 2021 transformation: learn more
  • Network function etc.

Become a member

Facebook AI can copy photo text styles from one word Facebook AI can copy photo text styles from one word

Back to top button