Facebook announced that they developed the Automatic Alternative Text (AAT) system so that the visually impaired could perceive the photos on the platform more easily.

Since 2016, Facebook has been trying to make photos more accessible to visually impaired people by using alternative text. The company’s Automatic Alternative Text (AAT) system uses a synthetic voice that allows the visually impaired to understand images in the Facebook feed using object recognition. Facebook announced that they made the system 10 times more reliable and perceptible by improving the artificial intelligence used in the system.

More understandable images with improved AAT
The company has expanded the existing object and concept categories and added new ones. The AAT can now recognize landmarks, activities and animal species. In this way, the system can provide more detailed explanations.

In addition, the position of objects and the relative size of the elements in a photograph can be included in the descriptions. So an explanation can now be heard as “The image of five people, two in the middle and three scattered around the edges” rather than “There are five people in the picture”. The system will be able to detect and highlight which element is the primary object in the described image, based on size and positioning.

To achieve this, Facebook researchers used an artificial intelligence model trained on poorly controlled data using billions of public Instagram views and hashtags. They redesigned machine learning models as a starting point to train new tasks called learning transfer.

Since Facebook mentions possibilities in its statements, the 100% accuracy of the system seems open to discussion. The company said that descriptions are available in 45 different languages ​​because they are told in simple words. Although the system cannot convey all the details of an image, it will allow visually impaired people to understand the main theme of that image more easily.

LEAVE A REPLY

Please enter your comment!
Please enter your name here