Google has revealed an artificial intelligence (AI) capability that can enhance an eight-pixel square blurred image by 16 times, restoring the lost data to bring clarity to the elements in the
photo.
Research released earlier this month describes the technology as being able to hallucinate the missing details and pixels
to sharpen the image. Similar to other neural networks, the technology was trained by showing millions of images. In this case, researchers used a database with 3 million training images, so the
technology learned faces and shapes. The next phase required the technology to focus on comparing 8x8 pixel images with those that could possibly be smaller versions of 32x32 images.
The
technology managed to get researchers "most of the way" toward recreating a higher-resolution image, but the outputs are blurry. So the researchers combined two different networks to achieve success.
The two networks managed to redraw the image.
Images have become a major focus for tech companies. Earlier this month, Facebook reported training a neural network with tens of millions of
photos and use artificial intelligence to identify the content in photos across its social site. The move makes it easier for users to search for images, even those that are not in their Friend
network, as long as the photos are public.
The model matches search keywords to images in the photos with some degree of probability. The technology originated from helping the blind to see an
image. Facebook's computer vision technology, which the company revealed that it had worked on in August 2016. At the time engineers in the
Facebook AI Research (FAIR) group called the technique "real-time classification," a tool to help people search and identify objects in images.