Advertisement by Easy Branches Network Contribute your contents

New Neural Network Tech Boosts Resolution Of Any Photo

By: forbes.com 3 months ago
New Neural Network Tech Boosts Resolution Of Any Photo

An almost spookily accurate way of restoring high-quality images from damaged originals using neural networks.

News New Neural Network Tech Boosts Resolution Of Any Photo

About New Neural Network Tech Boosts Resolution Of Any Photo

Online New Neural Network Tech Boosts Resolution Of Any Photo
How New Neural Network Tech Boosts Resolution Of Any Photo

A group of researchers from Russia’s Skoltech Institute and Oxford University in the UK has developed an almost spookily accurate way of restoring high-quality images from damaged originals using neural networks.

The system, developed by Dmitry Ulyanov, Andrea Vedaldi and Victor Lempitsky, is able to automatically enhance picture resolution, remove image noise and invisibly plug ‘holes’ in damaged photos with striking results.

Dmitry Ulyanov, Andrea Vedaldi, Victor Lempitsky

Deep Image Prior removing text from an image by intelligent inpainting

Named ‘Deep Image Prior’, the method differs significantly from existing techniques such as those used by Google to enhance low-resolution images on Google+ in that the neural network needs no prior training.

Existing methods typically require a large set of example images from which the neural network can ‘learn’ before it is able to apply this knowledge to new input. However Deep Image Prior skips this stage entirely, generating restored output files from a single input image without prior training.

The results often come very close to the best training-based algorithms and are a huge step up from traditional learning-free methods.

Deep Image Prior can, therefore, be used to improve image quality in a range of scenarios without the need for computationally intensive pre-training or human intervention, whether you want to enlarge a small image, remove noise or repair treasured damaged photos.

Here Deep Image Prior effectively removes the compression artefacts from a highly compressed JPEG image.

Dmitry Ulyanov, Andrea Vedaldi, Victor Lempitsky

JPEG Artifacts Removal

Here we see Deep Image Prior attempting to reconstruct an image from 50% of the original pixels.

Dmitry Ulyanov, Andrea Vedaldi, Victor Lempitsky

Deep Image Prior image restoration.

As a further demonstration, the team shows how Deep Image Prior can combine a pair of images, one taken with flash, the other without. This results in a final image with the natural appearance of the non-flash image, but the clarity and detail only present in the flash version. This process has been implemented before, notably in some Nokia Lumia smartphones, but Deep Image Prior improves significantly on the quality of the final result.

Dmitry Ulyanov, Andrea Vedaldi, Victor Lempitsky

Deep Image Prior flash / no flash combination

There are still significant limitations to Deep Image Prior’s approach. Because it works using only information from within the original image it can’t, unlike pre-trained neural networks, bring in any external knowledge to help it fill in any blanks.

As Dmitry Ulyanov explains, “The obvious failure case would be anything related to semantic inpainting, e.g. inpaint 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.”

For more details including interactive samples, see Deep Image Prior and this Reddit post from Dmitry Ulyanov.

___

More On Forbes

Samsung's Radical New Smartphone Camera Could Be Headed For Galaxy S9

iPhone 8, iPhone 8 Plus Tests Reveal Apple's Biggest Upgrade

Google's Pixel 2 Secret Weapon Is 5x Faster Than iPhone X

Google's Pixel 2 Has Big Win Over Apple's New iPhones

GoPro Hero6 Black Vs. Hero5 Black: What's The Difference?

GoPro Hero5 Black Vs Hero5 Session: What's The Difference?

Google's Brilliant iPhone App Sucks On Android

Follow Paul on Twitter, Instagram and Google+

NeuralNetworkTechBoostsResolutionPhoto