Let’s Enhance uses a neural network to upscale the images but unlike other attempts, the system first identifies the type of image, such as a portrait or landscape, then uses a separate network to enhance the image using some of the anomalies of the image type.
Along with using different networks for different types of images, the program uses three different layers to produce the final image. The first layer is the “Anti-JPEG” filter, dedicated to just removing the artifacts that are created. The second, called the “Boring” filter, does the actual upscaling while also working to try to preserve the details in the original shot.
The third layer is the one that makes the image look like it was shot at that original resolution in the first place, aptly called the “Magic” filter. While the first two layers work with what’s already there, the Magic filter uses the data from the photos in the database to make-up new details to fill into the larger file. This final layer might just be why the program tends to get better results than earlier attempts.
While the program is creating better upscaling results, that final layer could prevent the program from turning the fictional scenes in CSI from becoming reality, (you know, those scenes where the characters turn a grainy security camera photo into a high-resolution one). That final layer adds details that were not actually in the original photograph and a program using made-up details isn’t likely to stick in court. Besides just making photos better, upscaling, as Google research suggests, could help bring better pinch-to-zoom capabilities to smartphones.
The program was designed by Alex Savsunenko, who has a Ph.D. in chemistry, and Vladislav Pranskevicius, a former chief technology officer. The developers said that, as a neural network, the program will continue to improve with use and that the program currently works best with landscapes and wildlife.
So can artificial intelligence really create a better upscale photo? The program is free to try out at letsenhance.io.
