Researchers at Duke University have developed artificial intelligence capable of reconstructing a realistic portrait in very high definition from a poor quality image. The technology dubbed PULSE for Self-Supervised Photo Upsampling via Latent Space Exploration of Generative Models (oversampling of self-supervised photos via the exploration in latent space of generative models, in French) makes it possible to considerably improve the resolution of an image. According to the researchers, their AI can multiply the quality of a photo by 64.
The researchers behind the tool demonstrate the full power of their AI from their own portrait. They took a picture of themselves, reduced the resolution of their portrait and submitted it to the AI, which managed to reconstruct realistic portraits in high resolution. The process is particularly impressive as the result is realistic. There is still a small difference between the original and the production of AI. Currently, no program is able to reconstruct an entire image from incomplete data. The work carried out by this AI is therefore considered a major advance in the field.
Dual AI
The PULSE tool is actually based on a preexisting technology, that of generative opposing networks. These are algorithms capable of generating particularly realistic portraits. Two AIs are put in competition to generate these realistic images. The first AI is the generator, it generates a sample of images, including portraits. The second has the role of discriminator and is responsible for detecting whether the sample is real or the result of the generator. At each failure, the generator will try to improve his portrait to deceive his opponent which, in the end, gives a very realistic result.
The PULSE uses this technology to generate a realistic portrait from the blurred photo and constantly compares it so that the result looks as much as possible. "Never had super-resolution images been created with this resolution in such detail," enthusiastically one of the researchers.
The interest of this kind of tools is currently limited to the artistic framework or to research, but one can imagine that in the more or less near future, it will be integrated into photo editing software. It could indeed serve to improve portraits very simply or anonymize passers-by on a photo taken in a public place, for example.
High-Tech
Sony and Microsoft jointly develop image sensors based on artificial intelligence
High-Tech
A search engine allows you to find all the photos of a person on the Internet
- Art
- Sciences
- Artificial intelligence
- Technology
- New technologies
- Photoshop
- Photo
- High-Tech