Artificial Intelligence Allows Improving The Quality Of 19th Century Video Clip, Here's How!

In 1896, a silent film of 50 seconds, L’Arrivee d’un train en gare de La Ciotat, was released and many of the viewers ran away thinking the train in the film would come out of the screen and run over them. Even though it was a blurry and low-quality film, still people perceived it to be real.

Eyeballs of them would have popped out if they had seen the new AI-enhanced version, released by Denis Shiryaev with better quality.

Publicly available enhancement programs, DAIN and Gigapixel AI of Topaz Labs were used by Denis Shiryaev to covert the low-quality film into a 4K 60FPS clip. The proprietary interpolation algorithm is used by Gigapixel AI to analyze and recognize the details and structure of an image and then give it a final look.

Topaz has trained AI to sharpen and clear images that even after enlarging it by 600 percent, remains clear. Whereas, DAIN add frames in between the keyframes of the video clip, the same as the motion smoothing feature on 4K TVs. Enough frames are added by DAIN to increase the rate of the film to 60 FPS.

Since the first high definition television in the market, these have played a vital role in broadcast entertainment and uplifting the technology. The resolution of HD televisions is six times the resolution of SD. Standard definition televisions have a 720x480 resolution, 345,600 pixels in total whereas high definition televisions display 1920x1080 or at 2,073,600 pixels. On the other hand, 4K has a resolution of 3840x2160, which means 8,294,400 pixels.

To enlarge the HD image to fit on a 4K screen, 6 million pixels need to be added. The interpolation process comes when the upscaler look where to fit an extra pixel display. It analyses what the pixel should display, depending upon what the nearby pixels are showing. It can be measured in several different ways.

The “Nearest neighbor” method is used in which blank pixels are filled with the same color as the nearest pixel. It gives effective results but pixelated to some extent. Bilinear interpolation involves more processing, allowing the TV to identify every blank pixel based on the nearest neighbor pixel, creates gradient and sharpens the image.

Whereas, Bicubic interpolation analyze the nearest 16 neighbors, giving accurate colors to image but makes it blur.

Deep convolutional neural networks are used by programs like DAIN that analyze and place the filler images in the existing frames of video clips.

The results of images may not be as impressive as one of the video producer Chris Schodt noticed several visual artifacts. According to him, it looks better in YouTube-sized pieces but playing on the large screen shows the edges of objects in an image are not defined and look a little apart.

Though there are some deficiencies but techniques used by Shiryaev can be beneficial and could lead to tempting opportunities.

Read next: Google-owned Jigsaw comes up with a tool to spot fake and tampered images
Previous Post Next Post