Google announces "RAISR" technology using machine learning to improve the image quality of low-resolution images at high speed

Every day on the web, countless photos are shared, but the resolution of uploaded photos varies depending on the shooting equipment, while low resolution images are sometimes stretched and jaggy covered with jaggies. While high-resolution displays are becoming increasingly popular for both stationary and mobile devices, "low-resolution images that are easy to see" are becoming necessary, but new technologies using machine learning "RAISR(Rapid and Accurate Image Super-Resolution) "enables you to generate" high quality version of low resolution image "in real time.

Research Blog: Enhance! RAISR Sharp Images with Machine Learning
https://research.googleblog.com/2016/11/enhance-raisr-sharp-images-with-machine.html

[1606.01299] RAISR: Rapid and Accurate Image Super Resolution
https://arxiv.org/abs/1606.01299

"RAISR" announced by Google Research is based on machine learning,Super resolution technologyIt is a new technology that generates 10 times to 100 times faster "high-quality version of low-resolution video", it can be executed in real time as a general mobile device.

RAISR filters are trained by two methods. The first is a set of "low resolution version and high resolution version of the same picture", the second is to learn from "upsampled" images, which is a conventional technique to improve the image quality of low resolution images It is a method. Below is the original low resolution image and the image upsampled by the conventional technology are arranged side by side, the upsampled image has many pixels and it is easy to see, but you can see that the outline is blurred. By using a set of "Upsampling version and High Resolution version of the same picture" for learning, RAISR can generate "high quality version of low resolution video" quicker and more beautiful than conventional technology.


From the left, the following 3 pictures show the low resolution original photo, the photos corrected by upsampling, and the photos outputted by RAISR are lined up, so you can see that the RAISR version is finished sharper than the upsampled version .


Google Research also publishes several comparative samples. Below is the original on the upper side, RAISR on the lower side, and RAISR processing clearly shows the elderly's eyebrows and wrinkles.


Here you can see that the left is the original, the right is RAISR, the boundary between the outline of the horse's statue and the background is clear.


As this technology progresses further, it is possible to improve the performance of "pinch out" of the touch screen terminal and to capture, store and transmit the images stored in the mobile terminal and the cloud in a state that it is easy to see with low resolution It is said that.