'Restoring' Old BBC Shows with CNNs

Deep Learning
1

Hey everybody! So a few weeks ago, I got inspired by the odd effect of Jeremy’s super-resolution demo where it corrected the colour balance of an image and decided to see if I could create a neural net architecture to recolour some old BBC shows where the original recording has been lost, but it still exists in a b/w format. Early days, but my initial experiments seem positive:

http://snappishproductions.com/blog/2017/07/04/the-colour-problem---neural-networks-the-bbc-and-the-rag-and-bone-man.html

(at the moment, I’ve ended up with a partial U-Net architecture)

Now that my shiny 1080Ti has arrived, I’ll be gearing up for a run on the full dataset probably by the end of the month, and I’ll stick up the full code after I’ve cleaned it up some ;).

6 Likes
2

Creative project. Good write up. Happy to see you had some success.

Nicely done. I look forward to more.

3

Thanks for posting! This is some really, really cool stuff. Will be looking forward to following your progress. :grinning:

4

https://hackernoon.com/remastering-classic-films-in-tensorflow-with-pix2pix-f4d551fa0503

5

Thanks, all! @pietz - I had actually been wondering this week whether it might also be a good fit for getting to grips with WGANs, so thanks for posting that!

6

Ryan Dahl also has a great write up on some of his work on colorization:
http://tinyclouds.org/colorize/

7

Also somewhat related, I’ve spent some time on subpixel convolution, and have found it to be quite fast for upsampling.

Writeup: