Microsoft Bing has announced a new AI technology that will bring a 4K graphics experience to websites through Microsoft Edge, automatically enhancing website images. The technology, called Turing Image Super-Resolution, allows images to be displayed at high resolution, no matter how bad the original image is.
This new technology was developed by Microsoft Corporation Turing project artificial intelligence development team.
Used in Bing Maps
This new technology is already being used in Bing Maps to improve the quality of its satellite aerial imagery.
Below is a comparison of aerial images of Google’s headquarters in Mountain View, California.
The screenshot from Bing Maps is on the left, and the corresponding image from Google Maps is on the right:
Bing Maps vs. Google Maps
How Microsoft Builds Technology
There are four important insights that lead to the success of this model.
- human evaluator
- noise modeling
- Perception and GAN loss
- Visual Transformers: Enhancement and Scaling
human evaluator
Microsoft realized that the metrics used to measure the success of image-related models were inconsistent with human visual perception. So they created a side-by-side visual comparison tool that uses human evaluators to help assess the success of the models.
noise modeling
Microsoft took the approach of starting with high-quality images, then degrading them by adding noise to them, and then teaching the model to bring the images back to their original high-quality state.
Perception and GAN loss
This is part of an effort to make the results consistent with human vision.
Microsoft announced:
“…we found that optimizing our model using only the pixel loss between the output image and the ground truth image was not sufficient to produce the best output consistent with human eye perception.
In response, we also introduce perceptual loss and GAN loss, and adjust the best weighted combination of the three losses as the objective function. “
Vision Transformers
Microsoft leveraged the power of Transformers used in the language model, focusing on enhancement and scaling.
This means enhancing the image and focusing on enlarging it, which is a difficult thing to do.
Usually, scaling down an image is easy. But taking a small image and zooming in on it usually amplifies the low-resolution artifacts of the original image.
So what the researchers did was create a system that could compute and “recover” missing image data from lower resolution images and upscale it to higher resolutions.
Microsoft calls the process of zooming in on an image DeepZoom.
The Verge: A 4K TV with a Web Browser
Microsoft envisions this new AI capability as a way to bring a 4K visual experience to web surfing and enhance videoconferencing and family photos uploaded to the web.
The technology is already available in an experimental version of Edge called fringe canary.
The new features will roll out to mainstream versions of the Edge browser in the coming months.
