Computers were able to find human faces from simple brushstrokes

Using data from 17,000 faces on the Internet, scientists have taught AI to accurately detect faces thanks to basic brushstrokes.

Using data from 17,000 faces on the Internet, scientists have taught AI to accurately detect faces thanks to basic brushstrokes.

According to Futurism, a research team at the Beijing Academy of Sciences, China and the University of Hong Kong has successfully developed an algorithm that turns a sketch with a basic drawing into a real-life portrait with all the details above. face.

The researchers said they used the faces of 17,000 celebrities on the Internet as data sources to teach AI.

If we continue to update the face data source for AI, this algorithm may be used in the investigation of suspects, or in the animation industry, video games .

Picture 1 of Computers were able to find human faces from simple brushstrokes

Breaking into pixels and joining them into long lines is how the team trains AI. (Photo: arxiv).

To train the AI ​​system, scientists need a database big enough to get started. They decided to use the stock photos of celebrities on the Internet first, then continued to update faces in China.

The team identified a face that would consist of five key areas: two eyes, a nose, a mouth and a face circle with hair.

Each area of ​​the face is digitized into pixels. Multiple created pixels will form different seams. Using this approach, when there are seams drawing faces, AI will find the correct face back by pixel retrieval.

Currently, because the data the research team has are all white people from countries like the United States and the United Kingdom, the images that AI draws are all white results. Color factors such as skin color, eye color, hair color, and tattoo will be further developed by the team.

"It is not easy to gather a database of users' faces, we will focus on working to teach AI to distinguish skin colors in the near future," said Hongbo Fu, a member of the research team. about the future of the project.

Update 18 June 2020
« PREV
NEXT »
Category

Technology

Life

Discover science

Medicine - Health

Event

Entertainment