Facebook researchers have developed the first self-supervised artificial intelligence model that can transcribe text in handwritten scenes and in the real world using just one example of a word.
The past few years have shown that faces, voice and lip movements can be copied using artificial intelligence.
The company notes that while most AI systems can copy and replace text for well-defined, specialized tasks, TextStyleBrush is different because it can reproduce text in both handwriting and real-world scenes.
Doing so is more difficult for the AI model because of the different text options and nuances.
And Facebook indicated that this means understanding the unlimited text patterns of different typography and font and also different transformations, such as rotation, curved text, and distortions that occur between paper and pen when writing by hand.
According to Facebook, TextStyleBrush is the first self-supervised artificial intelligence model that can copy and replace text in handwritten scenes and in the real world using just one example of a word from an image.
The company said it showed high accuracy in automated tests and user studies of any type of text. Adding that she intends to submit the work to a scientific journal for judgment.
The AI model works in a similar way to the pattern brush tools used in word processors.
In this case, the method is applied to the aesthetics of text in images. However, Facebook researchers using TextStyleBrush decided to abandon the traditional methods. Which involves defining specific criteria such as writing style or oversight of the target style.
“We’re taking a more comprehensive training approach, separating text image content from all aspects of its appearance in the entire word box,” Facebook said.
She added: “The general appearance representation can then be applied as a starting point without retraining on new source pattern samples.”
However, TextStyleBrush is not without limitations. The company said: These restrictions include working with text written with metal objects or text with letters of different colors.
Facebook acknowledges that deep fake text attacks are a problem. By publishing her research publicly, she said, she hopes to encourage new research to prevent these attacks.
The company says that if AI researchers can build this technology first, they can learn how to better detect this new type of deepfake and build robust systems to combat it.