Image recognition is one of the most efficient AI algorithms available, and will likely be used to improve all types of media in the future. The dominant image-recognition algorithm today is Google’s Deep Dream. It has an error rate around 10% for identifying images that are not obviously a dog or car, making it over 100x faster than humans.”
The “microsoft edge web browser” is a new browser that will improve all images on the internet. It uses AI to do this and it is also faster than any other browser.
Microsoft Bing introduced a new AI technology that would automatically enhance website photos and offer 4K visual experience to webpages using Microsoft Edge. The Image Super-Resolution by Turing technology allows pictures to be shown at a high resolution regardless of how terrible the source image is.
Microsoft’s Project Turing AI development team created the new technology.
Bing Maps already uses it.
Bing Maps is already using the new technology to improve the quality of its satellite aerial photos.
Below is a comparison of aerial photography of Google’s Mountain View, California headquarters.
On the left, a screenshot from Bing Maps, and on the right, a screenshot from Google Maps:
Google Maps vs. Bing Maps
Microsoft’s Technology Development Process
There were four key observations that contributed to the model’s success.
- Human Judges
- Noise Simulation
- GAN and Perceptual Loss
- Enhance and Zoom with Vision Transformers
Microsoft realized that metrics used to measure success of image-related models didn’t align with human visual perception. So they created a side-by-side visual comparison tool that used Human Judges to help evaluate the success of the model.
Microsoft used the method of beginning with high-quality photos, degrading them by introducing noise, and then training the model to restore the image to its former high-quality condition.
GAN and Perceptual Loss
This was done as part of the attempt to make the outcomes more human-like.
According to Microsoft’s announcement:
“… we discovered that refining our models merely based on pixel loss between output and ground truth pictures was insufficient to yield the best output that matched human perception.”
In response, we also introduced GAN and Perceptual Loss and tuned an optimal weighted combination of the three losses as an objective function.”
Microsoft focused on enhance and zoom utilizing Transformers, which were previously employed in language models.
That implies improving the picture while simultaneously concentrating on scaling it up, which is a challenging task.
Shrinking a picture is usually simple. Taking a tiny picture and scaling it large, on the other hand, usually magnifies the original image’s poor quality artifacts.
So the researchers devised a method for calculating and “recovering” missing picture data from a lower-quality image and bringing it up to a higher resolution.
DeepZoom is the name Microsoft gives to the technique of scaling a picture up.
Edge: Web Browsers’ 4K TV
This new AI capability is envisioned by Microsoft as a method to bring a 4K visual experience to online browsing, as well as improving video meetings and family images posted to the web.
The technology is already available in Edge Canary, an experimental version of Edge.
Over the next several months, the new capability will be rolled out to the mainstream version of the Edge browser.
Read Microsoft’s Press Release
Image Super-Resolution by Turing
- microsoft edge image quality
- ai generated images