Facebook developed an AI that can identify offensive memes

Now while sharing photos/videos on Facebook, you should be more aware because Facebook is developing an AI that can understand the text on image and videos. This new AI, called Rosetta, which understand the content written on image/video and detect whether it violate Facebook’s policy or not.

On Facebook, millions of photos shared which contains text in the various form like a meme. Facebook’s target is to identify such images which contain inappropriate text, which makes people uncomfortable.


This AI follows two steps to extract the text from the image: detection and recognition.

  • The first step, it detects the rectangular regions of the text.
  • The second step, it recognizes the text and understands the meaning.

This AI uses a convolutional neural network (CNN) to recognize and transcribe the word in the region.

It is quite different from the Optical Character Recognition (OCR) system. OCR only recognizes the characters but don’t understand the context. They have built and deployed a large-scale machine learning system to solve the problem of understanding the text.

Rosetta works daily and in real time with the content that is shared on Facebook and Instagram, analyzing the text of one billion images and video frames. An interesting detail is that you can perform this dynamic in different languages and dialects.

It currently supports different languages and encodings such as, among others, Arabic and Hindi, in a unified model.



Be the first to comment