Google Brain develops technology to predict the original image from the image reduced to 8 × 8 pixels

Study on deep learningGoogle BrainIs a technique for estimating an original image from an image obtained by converting a high resolution image into 8 × 8 (64) pixels "Pixel Recursive Super ResolutionWe announced.

Pixel Recursive Super Resolution
(PDF file)https://arxiv.org/pdf/1702.00783.pdf

Google Brain super-resolution image tech makes "zoom, enhance!" Real | Ars Technica
https://arstechnica.com/information-technology/2017/02/google-brain-super-resolution-zoom-enhance/

The right edge of the image below is the original "source image", which is compressed to 8 × 8 pixels size is the leftmost "8 × 8 sample". The original image that Google Brain's deep learning technology expected from this 8 × 8 sample is "32 × 32 samples" in the middle column. You can see that the predicted image upscaled from the 8 × 8 sample to the 16 × resolution can reproduce the image closer to the source image from the less information of 8 × 8 samples.


Google Brain uses two neural network training to perform image prediction. One is a "conditioning network" that checks patterns and colors by comparing 8 × 8 samples with compressed data of similar high resolution images and anotherPixel CNN"Prior network" which adds high resolution details using high resolution. It is said that a predicted image is created by combining these two neural networks.


Four types of predicted images produced on the left edge by 8 × 8 samples and the remaining four from the neural network. It is possible to reproduce the image of the bedroom as well as the portrait.


Of course Pixel Recursive Super Resolution can not perfectly reproduce the original image, but by improving the technique of Pixel Recursive Super Resolution, you can try to improve the image which is common in the movie world "Try to zoom in (more)!" The scene of it may become a reality thing.

Let's Enhance (HD) - YouTube