Google AI brings TV crime cliché to life
Posted: 8 February 2017 | By Darcie Thompson-Fields
Google Brain have trained their AI systems to transform heavily pixelated, low-quality images into a clear photo of an object or person.
Computer scientists from Google’s artificial intelligence team have found that it’s possible to make even the blurriest of images recognisable. In a paper, Pixel Recursive Super Resolution, three researchers showed it was possible to not only enhance a picture’s resolution but fill in missing details as well.
The paper outlined how Google trained its system on small 8×8 pixel images of celebrity faces and photos of bedrooms. The system uses a combination of a conditioning neural network and a prior neural network to analyse the image to provide a higher resolution 32×32 pixel version.
“When some details do not exist in the source image, the challenge lies not only in ‘deblurring’ an image but also in generating new image details that appear plausible to a human observer,” the Google Brain researchers wrote in their paper.
To create a final de-blurred image, Google combines the results of two neural network processes. The first part tries to map out the source image against other high-resolution images, it downsizes the images and attempts to make a match.
The second part uses an implementation of PixelCNN to try and add realistic details to the source image.
The prior network ingests a large number of high-res images when the source image is upscaled, it tries to add new pixels that match what it knows about that class of image. As Ars Technica explained, if there’s a brown pixel towards the top of the image, the prior network might identify that as an eyebrow: so, when the image is scaled up, it might fill in the gaps with an eyebrow-shaped collection of brown pixels.
Currently, the AI creations are the machines best guesses rather than accurate portrayals. In the future, as the technology develops, it could be used to de-blur CCTV images.