Enlarging images with alpha channel

[xinntao]

Enlarging images with alpha channel

[Krischan74]

I have successfully converted all my RTCW/ET textures with ESRGAN and my own automated RTCWHQ method, see https://github.com/Krischan74/RTCWHQ

[Fraggoso]

I've found a new code that preserves alpha channel.
Credit goes to RWTema: https://pastebin.com/sb5PCKxT

I did a few testing and it works BUT it takes very long to process even on my GTX1080.
Normally a picture from 320x240 takes one second. With that new code it takes more than one Minute on the same picture.

[xinntao]

I think it is possible. After reading images, only process the color channels with the ESRGAN model and directly copy (or maybe bicubic upsampling) the original alpha channel. After that, concatenate the color channels with the alpha channels, obtaining the final result.

[xinntao]

I am not familiar with images with alpha channels. I do not know whether I understand correctly.

1. The simplest way may be to use the current model for RGB images and directly upsample the alpha channel. And finally combine RGB channels and Alpha Channels together.
   However, there are some problems when the edges of the alpha channel mismatch the RGB channels, I found an example for game images here .
2. So another improved method is to also upsample the alpha channel, treating it as a gray image using the current model. I am not sure what the final effects looks like, may have a try.
3. If having lots of training data with alpha channel images, we can fine-tune the current model with RGBA channels. (May also get the training pairs from other HD games, which will decrease the domain gap.)

If possible, could you @shogun1337 @Dilapidated provide some example images with RGBA channels? I may have a try with current models ;-)

[shogun1337]

OpenCV seems to have support for reading an alpha channel, theres a cv2.IMREAD_COLOR flag that i've changed to cv2.IMREAD_UNCHANGED but sadly seems it nothing changed

[Dilapidated]

Having this supporting alpha channel on output would be awesome. Probably more of a BasicSR question but I was curious if using images with alpha channels would cause issues currently with training as well?

[shogun1337]

@xinntao Sure, there are some texture samples with transparency from the "Return To Castle Wolfenstein" game
https://www.mediafire.com/file/8nu5gyy1asvq0mf/samples.rar/file

[Dilapidated]

@xinntao Thank you for looking into this deeper. I have included examples from two different games for variety/style purposes. Resident Evil 2 and New Super Mario Bros.
https://www.mediafire.com/file/f6twysc62h6kfpg/ALPHA_EXAMPLES.zip/file

[shogun1337]

The results are also pixelated, maybe did he used the nearest neighbour interpolation?

[Dilapidated]

@Fraggoso Thank you for sharing this. It's good to see progress being made on retaining the alpha. Sadly I get the same pixelated artifacts as Shogun1337 and it does seem to take quite a bit longer. Hopefully @xinntao can further refine the code to fix these issues.

[Fraggoso]

@shogun1337 Can you post your image source so I can try it as well?
Like I said, it worked on my end.

[xinntao]

I have tried it with this script https://pastebin.com/xUdTHff9
It seems that the results are also not satisfying.
The results can be downloaded from here.

To control the noise and details, I also use the interpolation between RRDB_ESRGAN_x4.pth and RRDB_PSNR_x4.pth.

[Fraggoso]

@xinntao The results, at least for me, were expected to look like they do.
Your script isn't working on my end though. Did you change anything? I receive an error on line 35 img.shape

[shogun1337]

@Fraggoso You can find this image in the pack i've posted ealier
(https://www.mediafire.com/file/8nu5gyy1asvq0mf/samples.rar/file)
@xinntao I think the results won't be better, even with extracting alpha channel from the original picture, bicubic upsizing in photoshop i'm getting very similar result:
(open this image on white background to see bicubic pixel artifacts with photoshop)

[Fraggoso]

@xinntao I receive an error if I run the script:
Traceback (most recent call last): File "test.py", line 35, in <module> H, W, C = img.shape
What am I missing?

[shogun1337]

@Fraggoso
Hmm it works good with python 3.5 x64 (for windows), opencv 3.4.1, numpy 1.14.5+mkl, torch 0.4.1
Have you tried updating dependencies?

[Moukrea]

Isn't it a way to simply ignore alpha channel and save the upscaled image without any alpha? That'd be good!

[Malkierian]

Has there been any progress with this? Upscaling a game like Metroid Prime Trilogy will take forever without alpha channel handling. I can't even figure out how to programmatically sort the PNGs that are dumped by which ones have alpha and which ones don't, so I can process the ones with transparency with waifu2x-caffe, so I'd have to go through 10k+ images one by one and manually sort them.

[shogun1337]

@Krischan74 incredible work, thank you very much, this can be used for every other games aswell

[Krischan74]

Thanks. I've created this for my "Enemy Territory Special Edition" (ETSE) which is still under construction / testing. It won't appear on Github but perhaps on moddb as it is LARGE (10GB!). At least the cool map launcher ETlaunch will be in my repository soon.

Oh and many thanks to the guys from ESRGAN without ETSE wouldn't be possible. Before I found your cool project I upscaled thousands of textures MANUALLY in Photoshop with a special template I created. That looked already great but the ESRGAN results are even more stunning :-)

[Fraggoso]

Is it possible to edit the code so that you don't have to check if an image has alpha in them or not?
I'm not a coder, sadly or I could try to do it myself but I really think it should be an easy fix to tell in test.py IF the image has alpha than proceed with X if not than proceed with Z?

[Robolightning]

If anyone else is interested in working with the alpha channel, you can look at my new repository, where I am creating a fork of the Real-ESRGAN with multi-channel support (essentially training with any number of channels, in particular 4-channel, with alpha channel support)
https://github.com/Robolightning/Multichannel-Real-ESRGAN