Colorizing and restoring old images with deep learning

Post by samstave » Sat, Nov 03, 2018, 9:11 PM UTC

dsfyu404ed wrote: post #8

iforgotpassword wrote: post #3

I've to admit I have no clue about machine learning, but what I notice is that this seems to have preferred colors for things that can actually have many different colors, most notably clothes. They're almost always this blueish slightly purple color here, even the samurai. Don't get me wrong, this is still awesome and I might try this on some old photos from my grandparents. I'm just wondering if and how one can pre…

The Seneca Native in 1908 example seems the most absurd to me. I know the software has no notion of a "fabric" or "clothing" but it's very rare for brown or beige things to fade to blue (or vise versa). In real life things when transition from brown/beige to another color that other color tends to be a red orange or yellow. I know from the known issue that it likes blue but it still seems very odd that it chose to fa…

Post by MayeulC » Mon, Nov 05, 2018, 12:27 PM UTC

citnaj wrote: post #102

MayeulC wrote: post #99

Earlier quoted context omitted.

Hey, thanks for replying! However, I feel like you glossed over the proposed workaround, which I feel is appropriate (though more complicated if you want to implement"defade"), and extremely easy to implement. I took a couple minutes to write an octave script that implement the workaround [1], it would have been even easier if both images had already been distinct files, and perfectly aligned. The basic idea here is…

Oh shit yeah that really does look good! Amazing really. Ok...I'm going to put these notes on the project board. Yes..I definitely glossed over the proposed workaround and I apologize. Thanks for this.

Post by tripzilch » Tue, Nov 06, 2018, 9:57 PM UTC

taeric wrote: post #79

bitL wrote: post #76

Earlier quoted context omitted.

It's Deep Learning, not much to do with any analytical model, it's not thinking like a human :-(. Recently even good NLP processing needs 24GB+ for training (won't fit into 16GB), a good quality colorizing (no spills, natural colors) could be expected to be as demanding. From the article: "BEEFY Graphics card. I'd really like to have more memory than the 11 GB in my GeForce 1080TI (11GB). You'll have a tough time wit…

I get that. I just have a hard time thinking that is "generalizing" the model, so much as making the model all encompassing.

Post by MayeulC » Thu, Nov 22, 2018, 10:59 AM UTC

citnaj wrote: post #106

MayeulC wrote: post #103

Earlier quoted context omitted.

No problem :) Although I would have made it a fully-fledged github issue, with a link in your board, instead of a text entry, to add supplementary material in the issue thread. Bonus: if you are only interested in chrominance, you can train your network to use YUV as an input instead, and output only UV. I suspect this might lead to substantial gains in the training time and network complexity.

Update: I got this working, and dude- it's so awesome in every way. This is the most substantial improvement I've seen yet. Most importantly- it massively reduces memory requirements. Thank you so much. I'll commit within a day or so and make sure to mention you, on Twitter.