Deep convolutional neural networks (CNN) have become a popular tool for image generation and restoration, with a few algorithms released in the last months of 2017 that achieve spectacular results when it comes to recovering and reconstructing corrupted or low-quality images.

One of the most impressive algorithms is called Deep Image Prior, developed by a team of Russian scientists.

Deep Image Prior stands apart from the crowd because it deviates from the norm. Instead of relying on large quantities of training data to determine the best way to deal with a degraded image, Deep Image Prior uses data from the degraded image itself to reconstruct the original photo.

Researchers say their algorithm was able to denoise images, remove text from photos, re-fill cropped images, remove pixelation caused by JPEG aliasing, and even enhance low-res images to a higher resolution with acceptable results.

The ability to re-fill altered images is very impressive and is the first time we see scientists recreate the "Content Aware Fill/Brush" features that Adobe added in Photoshop a few years back, a technology it kept secret and that no other image processing software maker was able to reproduce.

Deep Image Prior

Deep Image Prior

Deep Image Prior

Deep Image Prior

Deep Image Prior

But Deep Image Prior is just one of the great CNN research projects working with image processing that came to fruition this year. Another one is PixelNN.

Developed by three researchers from the Carnegie Mellon University, PixelNN can reconstruct fuzzy, pixelated, or incomplete images. The algorithm needs to be trained with loads of data, but it's more precise than similar projects, recreating images from highly corrupted data.


Another great algorithm is EnhanceNet-PAT that is specialized in upscale operations, taking low-res images and recreating a detailed high-res version.

While Deep Image Prior can also upscale images, its results are nowhere near the capabilities of EnhanceNet-PAT, which is far superior. But just like PixelNN, EnhanceNet-PAT needs to be trained with troves of sample training images before being deployed in the wild.

Researchers hope that EnhanceNet-PAT will eventually find its way into software for upsampling old movies to 4K quality, in restoring old family photos, or in improving low-res CCTV images in police investigations. It's results are quite promising as is.


Related Articles:

Microsoft Reveals a Unified AI-Powered Search Experience Across all Platforms

New Attack Recovers RSA Encryption Keys from EM Waves Within Seconds