Google Brain generates images based on pictures with a resolution of only 8×8 pixels
Neural networks have learned how to do the impossible. The Google Brain team is using a pair of neural networks to generate images based on pictures with a resolution of only 8×8 pixels. They synthesize realistic details while increasing the resolution of the original images.
The process and scientific work created by employees of the Google Brain research project is called Pixel Recursive Super Resolution. As they write in the white paper: AI synthesizes realistic details in images while increasing their resolution.
The process consists of two steps. First, the program takes an 8×8-pixel image, from which it looks for a similar, higher-resolution sample to produce a sketch of the portrait or scene being processed. The second network then adds resolution to the image by estimating pixel values based on other high-resolution images found on the Internet with a similar pixel arrangement.
The process is far from perfect, but the result is still impressive. In the image above, the left column shows the original 8×8 pixel images that the neural networks worked with; the second column shows the AI-generated images; the third column shows the faithful versions of the portraits, reduced to 32 pixels, from which the images for the first column were derived.
More information about the study and other samples of enlarged images can be found in a paper provided by Google Brain.
Scientists have created an algorithm that gives photos the style of paintings by great artists
Researchers from the University of the German city of Tübingen have described how their neural algorithm creates new artistic images by giving random photos the style of paintings by famous artists. The accomplishments were described in a recently published research paper entitled “A Neural Algorithm of Artistic Style.”
“We are introducing an artificial system based on a neural network that creates high perceptual quality artistic images. The system uses neural representations to separate and recombine the content and style of arbitrary images, providing a neural algorithm for creating artistic images.”
The algorithm drew on various famous paintings to create a new creation.
German scientists say that with this software they are trying to artificially recreate the human ability to create unique visual images “through a complex interaction between content and image style.”
Whether photo editing software based on this algorithm will appear is not yet reported.
Free online service improves photo quality by increasing resolution by 4 times
The secret of the software is the use of neural networks – computing systems organized on the principle of biological neural networks (human and animal brains). The current version of Let’s Enhance uses the neural networks most focused on landscape and portrait images to quadruple the quality of low-resolution images.
The program uses three layers to compose the final image. The first, the Anti-JPEG filter, removes JPEG compression artifacts such as edge loss, sharpness and tone. The second layer, the Boring filter, upscales (increases resolution), preserving the original details and edges. Using AI technology, the third and final layer, called the Magic filter, tries to identify the depicted object (based on previously uploaded photos in the database) and fill it with new details.
The results are surprisingly accurate, and the tool will continue to improve with further use.