We’ve seen some of the algorithms that can enhance low-quality photos. The researchers from Oxford University and the Skolkovo Institute of Science and Technology in Moscow have developed a new approach for restoring damaged or low-quality images. Instead of training the neural network with thousands of photos, their system called Deep Image Prior works everything out from the original image. And without any previous learning, it turns a pixelated or damaged photo into a hi-res one.
Dmitry Ulyanov, one of the co-authors of the research, explains that the “network kind of fills the corrupted regions with textures from nearby.” Instead of using the data from the datasets, Deep Image Prior redraws a blurry or damager picture until it gets it right. According to Interesting Engineering, some images turn out to be even better than the original input.
Despite the good overall results, Ulyanov admits that there are still failures in the redrawing attempts: “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.” The researchers mention some of the possible legal and ethical applications of this method. It can be used to filter out noise or upsample low-res photos, which photographers might find useful. Deep Image Prior can also be applied to the restoration of old and damaged photos, which museums and archives are likely to find useful. But there’s a downside to this neural network as well. Other than restoring damaged, blurry, grainy and otherwise corrupted photos, Deep Image Prior can also remove the text placed over the images. This raises copyright concerns, considering that the algorithm could make image theft easier than ever. Another question is whether apps like this could leave photo restorators without work. Sure, human brain and skill are still better than computers, so this certainly won’t happen today. But at some point in the future when the algorithms are perfected, the photo restorators’ jobs may become less wanted and less paid. I think we can definitely view discoveries like this from both sides. There’s no doubt there can be very useful, but also harmful if they get into the wrong hands. You can view more image examples here and read the full paper on this link. [Deep Image Prior via Interesting Engineering, Inverse]
