Researchers have developed a neural net that can turn grainy, out of focus pictures into super sharp snaps. The collaboration between the Oxford University and the Skolkovo Institute of Science and Technology in Moscow has led to the development of the neural network system they call Deep Image Prior. A neural network is best described by Dr. Robert Hecht-Nielsen who defines it as, “a computing system made up of a number of simple, highly interconnected processing elements, which process information by their dynamic state response to external inputs.” It uses information it is presented with to learn new skills and modes of processing. They are designed in a similar way to a human brain; networks are constructed from thousands of nodes that they use to make decisions about the data that is presented to them.
Network learns by doing rather than big data
Many neural networks learn by being fed large data sets that they use to train themselves in a particular task, but Deep Image Prior uses a different approach. Instead of using a large data set, the network was asked to redraw a blurry picture thousands of times until it got really, really good at it, and could create images that were even better than the original. The network uses the existing input to help it fill in the gaps of the missing or damaged parts. Dmitry Ulyanov, co-author of the research, describes the process: “[The] network kind of fills the corrupted regions with textures from nearby.” He does admit that there are times when the network fails in its attempt at redrawing. “The obvious failure case would be anything related to semantic inpainting, e.g. in-paint a region where you expect to be an eye — our method knows nothing about face semantics and will fill the corrupted region with some textures.”
Deep Image prior raises copyright concerns
Deep Image Prior actually got so good at ‘restoring’ images it was able to successfully remove watermarks placed over the top of the images. While this feature does raise concerns about copyright infringement, it isn't the first tool out there capable of such a task. The researchers are keen to point out, aside from the interesting applications the system can have for restoring photographs, it is a great example of how a neural network can function without the need for a large dataset as a base. The three researchers that developed the network, Andrea Vedaldi , Victor Lempitsky, and Dmitry Ulyanov, have published their code freely and made it available on GitHub. Deep Image Prior is just one of many new systems out there capable of improving the quality of images.
Museums and archives are likely to be very interested in the application of these systems to restore damaged and poor quality materials. As computing processing continually improves, perhaps domestic versions of the systems will become available for amateur photographers to also improve the quality of their work.