"EnhanceNet-PAT" that can convert low resolution photos of Gabi-gobi to high-resolution pictures appeared

As in the world of fiction such as movies, it is difficult for anyone who has edited or modified images to know that it is difficult to convert low resolution photos to high resolution ones. However, it is a German research institutionMax Planck Institute for Intelligent SystemsComputer scientists belonging to the company develop algorithms to convert low resolution photos into high definition pictures, and their performance is amazing.

[1612.07919] EnhanceNet: Single Image Super-Resolution Through Automated Texture Synthesis
https://arxiv.org/abs/1612.07919


AI method to upscale low-resolution images to high-resolution
https://techxplore.com/news/2017-10-small-pixel-perfect-large.html

New algorithm helps turn low-resolution images to include photos, 'CSI'-style
https://www.digitaltrends.com/cool-tech/algorithm-low-res-high-res/

The algorithm developed by the Max Planck Institute for Intelligent Systems is called "EnhanceNet-PAT", which uses artificial intelligence (AI) to create high-resolution images from low resolution images. According to Mehdi MS Sajjadi, one of developers of EnhanceNet-PAT, research to create high-resolution images from low-resolution images has been conducted so far.

Sajjadi said that it succeeded in creating a high resolution version of the image from the low resolution image by taking a different approach from the super resolution technology so far,Super resolution technologyThe issues related to the past have been studied for many years. Prior to this research, even the state-of-the-art super resolution technology had blurred or rough parts in the image. The reason for that is because we requested the neural network to realize what is impossible in a neural network such as "reconstruct the original picture so that it becomes a high-definition image". Therefore, it seems that some of the images created by the super resolution technology were blurred. Therefore, we took a different approach to create a high-definition texture with a neural network. In this method, the neural network sees the entire image, detects the region, creates a higher definition texture using semantic information, and creates a high resolution image. "

Using EnhanceNet-PAT, create a high resolution image (middle) from the low resolution image (left). The image at the far right is the original high resolution photo, you can see how the photos created by EnhanceNet - PAT are finished to be close to the original image.


In order to train the algorithm of EnhanceNet-PAT, the development team gave a huge dataset to the neural network and built knowledge of various textures. The neural network is a simple one that only looks at the downsampled version of the image, and derives from this how to make the low resolution image high resolution. When the network creates images, researchers compared EnhanceNet-PAT to be able to convert to finer images by fine-tuning the algorithm compared with the original high-resolution photographs.

EnhanceNet-PAT can be used in various ways, such as increasing the resolution of old movies, restoring old photographs, and making them high-definition when prolonging photos. It is also clear that when using the EnhanceNet-PAT algorithm's super resolution technology for other neural networks, it is possible to improve the accuracy of object detection in images. For example, if applied to a neural network that is used in Google's automatic driving car etc., it seems that pedestrian detection can be performed more accurately.