Upscale low res card images to high res using Deep Learning

[Huggybaby]

Sorry for the questions, but I'm not tuned in like I used to be...

How much time between Gatherer and HQ releases? HQ requires card acquisition and the time for that varies a lot but I don't know by how much lately, charlequin would know but since we're talking already I thought I should ask. lol

Also, these processed pics look really good but are smaller than yours and have less artifacts:
viewtopic.php?f=26&t=16678&start=120#p221247

Can you fill me in? Or Bog Wraith?

I guess what I'm asking is what's going on with the LQ Pics scene?

[austinio7116]

Looks like that other thread is using Waifu - a more generic anime upsampling tool that does a great job on art and gives a really smooth result, but less sharp than mine - it is trained on anime images rather than specifically magic cards, but is probably trained on a larger dataset for longer. I prefer my results at the resolution I used them at, but that could be a matter of taste.

I'm not sure what the lead time on HQ releases is.

[Huggybaby]

I love the way your pics look. But the thing that really draws my attention is at the bottom of the card where the credits and copyright are, that's where I have always immediately looked to assess processing quality. In the waifu everything looks clear even zoomed larger than yours but the Deep Learning has obvious artifacts.

I hope you're not offended by the observation because I don't mean to disrespect, or to waltz in here johnny come lately and tell you anything. Plus, I could be doing something wrong. NTM my ignorance.

[austinio7116]

It has been a problem to get that footer text looking good. Something about it seems to cause the network to really struggle.

With all the different parameters I tried, this current version is the best I have managed for the footer text. If I ever get time to revisit the model and training it would be an area to focus on.

[austinio7116]

Just for completeness, the Waifu algorithm is very similar, it also uses deep convolutional neural networks, just with a different model and different training data.

[Huggybaby]

Thank you. Do you mask out areas of the card so you can use different parameters for different parts?

[austinio7116]

Not currently, I leave the network to learn whatever it needs to learn, that is the joy of deep learning. Perhaps some hand holding would help, but it might make it less generic to different card frames etc.

[Bog Wraith]

Sorry for the questions, but I'm not tuned in like I used to be...

How much time between Gatherer and HQ releases? HQ requires card acquisition and the time for that varies a lot but I don't know by how much lately, charlequin would know but since we're talking already I thought I should ask. lol

Also, these processed pics look really good but are smaller than yours and have less artifacts:
viewtopic.php?f=26&t=16678&start=120#p221247

Can you fill me in? Or Bog Wraith?

I guess what I'm asking is what's going on with the LQ Pics scene?

The last set was one of the fastest turn arounds I can remember. I expect that this new set will also be available in HQ relatively soon after release. In the meantime, this upscale method gives us a really fine look to play with and I am most impressed with the software that does this. I agree that the bottom of the card's text is the weakest part of this process & I hope that austinio will be able to tweak it as we go forward, but this is such a wonderful step upward in quality that we can play with while we await the HQ release(s).

Nice to see you, my old friend. I trust you & yours are doing well & I wish you all the best in 2018!

[Huggybaby]

Thank you Bog Wraith. Yeah, we're still hanging on by our fingernails.
(Oh, and really enjoying this cold-a** weather you sent our way.)

Happy New Year brother!

[Bog Wraith]

Thank you Bog Wraith. Yeah, we're still hanging on by our fingernails.
(Oh, and really enjoying this cold-a** weather you sent our way.)

Happy New Year brother!

LOL!

Yeah, sorry about that one my brother. There are lizards falling out of the trees in Florida, FROZEN to death!

Climate change means just that, not that it is always getting hotter, which it is as the global temps continue to rise, but it means weather patterns are changing and out of the norm. This is just the beginning and things are going to get a lot worse, especially for the kids of tomorrow.

Stay warm my friend and think thoughts as I do, that pitchers & catchers report for Spring Training in a month or so.

Baseball is my salvation!

[Fizanko]

I can't believe i didn't spotted this thread before.
And the less i can say is that i'm very impressed by how good the processed pictures are considering they're coming from a low resoluton original.
Fantastic work !

[Bog Wraith]

The Commander Anthology set, or CMA, as you can see below, is only available in a very sub par scan AFAIK.

Not sure if your method can do anything with these as they are so blurry to begin with, but I'm posting this link so you can give it a go if you deem that they are even worth the effort.

All the images have the .full extension as that are used in Forge.

http://www.mediafire.com/file/438lbh53hdjduf2/CMA.zip

[austinio7116]

As those cards are lower than the 370x265 my network is trained on, I thought I'd try Waifu to upsample to 370x265 then put them through my network. Here are the results:

Low Res

Waifu

Waifu + SREZ

[austinio7116]

I don't think adding my network on top of Waifu helps - I'd suggest either using Waifu (which is a more generic upsampling network), or I'd have to re-train my network for these lower res images which I don't really have time for at the moment.

[austinio7116]