The AI that can make grainy images razor sharp

Spread the love
  • The new method uses neural networks to enhance low-quality images
  • It can create sharper results than traditional methods, to fill in the details
  • While the technology is common in science fiction, it’s remained a challenge

Cheyenne Macdonald For Dailymail.com

Researchers have unveiled a Blade Runner-style AI that can enhance pixelated images.

Dubbed EnhanceNet, the system relies on neural networks to boost the image quality, creating high-resolution images from a low-resolution input.

Stunning results shared in the study reveal how the new method can create sharper, more detailed results than traditional approaches, leading to photos that are nearly indistinguishable from the real thing.

Scroll down for video 

Dubbed EnhanceNet, the system relies on neural networks to boost the image quality, creating high-resolution images from a low-resolution input

HOW IT WORKS 

The system uses adversarial training, which pits the networks against each other in competing goals so they ultimately learn together. 

‘By using feed-forward fully convolutional neural networks in an adversarial training setting, we achieve a significant boost in image quality at high magnification ratios,’ the researchers explain in the study.

The system uses adversarial training, which pits the networks against each other in competing goals so they ultimately learn together.

In a paper published to arXiv, the researchers from the Max Planck Institute for Intelligent Systems explain that enhancing low-resolution images has long posed a challenge, despite its prevalence in science fiction.

‘The problem is inherently ill-posed as no unique solution exists: when downsampled, a large number of different HR images can give rise to the same LR image,’ the authors wrote.

‘For high magnification ratios, this one-to-many mapping problem becomes worse, rendering SISR [single image super-resolution] a highly intricate problem.

‘Despite considerable progress in both reconstruction accuracy and speed of SISR, current state-of-the-art methods are still far from image enhancers like the one operated by Harrison Ford alias Rick Deckard in the iconic Blade Runner movie from 1982.’

Much of the problem lies in recreating the textured regions, which often end up looking blurry and unnatural when enhanced.

Using a convolutional neural network, however, the researchers have revealed a way to overcome the issue.

The system uses adversarial training, which pits the networks against each other in competing goals so they ultimately learn together.

Examples from the study show just how good the new AI is, far surpassing the quality of images enhanced using a method known as peak signal-to-noise ratio (PSNR).

Eventually, this ‘could easily be put in commercial software such as Photoshop,’ lead author Mehdi S. M. Sajjadi told FastCoDesign.

Or, it could be incorporated ‘smartphones, enhancing the quality of images as you zoom into them to avoid blurriness.’

Stunning results shared in the study reveal how the new method can create sharper, more detailed results than traditional approaches, leading to photos that are nearly indistinguishable from the real thing

Stunning results shared in the study reveal how the new method can create sharper, more detailed results than traditional approaches, leading to photos that are nearly indistinguishable from the real thing

Stunning results shared in the study reveal how the new method can create sharper, more detailed results than traditional approaches, leading to photos that are nearly indistinguishable from the real thing

The details in the reconstructed faces are not necessarily true to the original, as much of the information is lost in the pixilation process, the researcher noted. Still, it could be used to help identify people – potentially posing privacy risks

The details in the reconstructed faces are not necessarily true to the original, as much of the information is lost in the pixilation process, the researcher noted. Still, it could be used to help identify people – potentially posing privacy risks

The details in the reconstructed faces are not necessarily true to the original, as much of the information is lost in the pixilation process, the researcher noted. Still, it could be used to help identify people – potentially posing privacy risks

The technique could be used by law enforcement, for example, for license plate recognition, allowing officials to obtain a plate number far faster, according to FastCoDesign.

And, it could be used to un-pixelate faces that have been anonymized.

The details in the reconstructed faces are not necessarily true to the original, however, as much of the information is lost in the pixelation process, the researcher noted.

Still, it could be used to help identify people – potentially posing privacy risks. 

In a paper published to arXiv, the researchers explain that enhancing low-resolution images has long posed a challenge, despite its prevalence in science fiction

In a paper published to arXiv, the researchers explain that enhancing low-resolution images has long posed a challenge, despite its prevalence in science fiction

In a paper published to arXiv, the researchers explain that enhancing low-resolution images has long posed a challenge, despite its prevalence in science fiction