Anti-Aliasing: Best Solution for Aliasing in Low Resolution RAW video Upscaling

Started by apefos, June 07, 2013, 06:21:50 AM

Previous topic - Next topic

0 Members and 1 Guest are viewing this topic.

apefos

In my tests with T3i 600D shooting 768 x 423 @37,7fps and 960 x 544 @ 24fps and upscaling to 1920 x 1080 I found the best solution for removing aliasing:

Red Giant Magic Bullet Instant HD 1.2 (99 usd plug-in)

Best settings for 768x432 and 960x544 resolutions:

output size: custom
1920 x **** (do your aspect ratio calculations) for 960x544 it is 1920x1088, for 768x432 it is 1920x1080
lock aspect ratio: off
filter type: better (important, "better" is even better than "best")
sharpness: 13 (more than this hurts the image)
quality: 8 (more than this hurts the image)
antialiasing: 10 (the maximum possible)

these settings I found after try everything with careful comparisons

you will not believe your eyes.

1%


a1ex

Aliasing can't be removed in post without breaking the Nyquist theorem ;)

(which is on the same level as breaking physics laws)

apefos

@1%, @a1ex,

Hi guys, did you try it? the results are amazing, it deserves a try...

I did a look at the theorem, but my maths is not enough to understand... maybe the Red Giant did a break on it?

Please, can you share how to make the upscaling in ACR to get good results? It will be useful for lots of people...

thanks!

apefos

I did another test and realized that apply two instances of the filter makes things even better:

first instant hd filter settings:
custom
1280x720
better
00
00
10

second instant hd settings
custom
1920x**** (height depends on what resolution you are shooting, for 960x544 = 1920 x 1088, for 768x432 = 1920x1080)
better
08
08
10

apefos

I did some more tests comparing apply the instant hd once or twice and I realized that twice adds some artifacts to the image, so apply once will be better. apply it once there is no artifacts.

adijiwa


apefos

There is another plug-in called ReSizer which comes togheter in the Instant HD package. It works great also, and delivers slightly different results, and it needs a small amount off sharpen filter also. you can compare both and decide what you prefer... for my taste the instanthd is better than resizer for anti-aliasing results.

here is a crop from a small part of the image for comparison, (240x150 pixels section from the 1920x1080 image). it is not perfect, but it is much better:



Audionut

I had good results with this back in the day.

If you're handy with avisynth, you can try changing the resizing algorithms in the script for better results.

Apply a good sharpener at the end of your script to break the laws of physics :P

Be prepared to significantly increase your processing time.  But hey, quality comes at a cost ;)

apefos

Improving results for 960x544@24p...

All the settings in Instant HD in my previous posts was found for ACR sharpening in AMOUNT=25, RADIUS=1, DETAIL=25

I realized if all sharpening in ACR is set to 0 (zero), Instant HD makes no difference because there is no significant aliasing in the image to correct, but this way, the overall looking is too much soft...

So I started to find the best balance between ACR sharpening and Instant HD correction...

Best results for 960x544 upscaling to 1920x1088 I found was:
setting ACR sharpening AMOUNT=30, RADIUS=1, DETAIL=10
setting Instant HD to sharpening 13, quality 08, anti-aliasing=10, mode=better
more or less the same thing as before, no other sharpen was used
if ACR is the same, this also works in Instant HD: sharpening=00, quality=08, anti-aliasing=06, mode=better

Audionut

Quote from: apefos on June 07, 2013, 07:10:30 PM
I realized if all sharpening in ACR is set to 0 (zero), Instant HD makes no difference because there is no significant aliasing in the image to correct

I am at an absolute loss to understand why you would increase aliasing just to later find the best way to remove it  :o

Find a better resizing algorithm that doesn't introduce aliasing in the first place ;)

apefos

It is so simple:

If you do not do sharpening in the image, it will be too much soft, no good to look at.

But there is a problem: every sharpening filter, effect or plugin from any software will increase aliasing. So this is the challenge: to do some amount of sharpening without hurt the image with aliasing.

Upscaling will magnify the image, so imperfections will be more easy to see. So to do sharpening and upscaling at the same time is something very difficult to do considering we want good results.

In my tests I found the Instant HD upscaling is the best because it corrects the aliasing found in low resolution video.

In my tests I found ACR sharpening is the best results to my eye, maybe because it is applied in the raw file using all the bits before converting.

So, I did tests in three way of doing the sharpening:
1- in ACR before Instant HD upscaling
2- inside Instant HD at the moment it does the upscaling (it has a sharpen slider)
3- using sharpen effect after Instant HD upscaling

The best results was from option 1: apply sharpening in ACR before upscaling and do upscaling in timeline using Instant HD.
Maybe this was the best result because sharpening in ACR uses all the power from raw file and doing upscaling with Instant HD uses the power of it's anti-aliasing which is very good.

As you can see I did try all the options and this was the best results. Using sharpening slider  inside Instant HD is optional, it is a matter of taste because sharpening was already done in ACR, so the different settings in Instant HD in my previous post is for you to choose the best for your eyes.

jchristman

I just came across this optical filter solution for moire. I'm intrigued:

http://recordingsofnature.wordpress.com/2012/12/04/cyl-filter-moire-and-aliasing-filter-for-hd-dslr-video/

This one is new to me. I've seen the  Mosaic Engineering AA filters, but it's my understanding that they won't work properly at resolutions other than 1080P.

Jason

Audionut

Apefos.  Could you be so kind as to upload a source image (dng please),  and a image that you have upsized and sharpend?

danielcreed

am i correct in assuming that in 10x mode you don't get any moire or aliasing since there is no line skipping ?

i understand that it is using just a cropped portion of the sensor.

N/A

So is it just the footage I have or does ACR process 2003 handle upscaling MUCH better? Someone wanna test it out?
7D. 600D. Rokinon 35 cine. Sigma 30 1.4
Audio and video recording/production, Random Photography
Want to help with the latest development but don't know how to compile?

ashtrai

Quote from: Audionut on June 07, 2013, 09:05:29 PM
Apefos.  Could you be so kind as to upload a source image (dng please),  and a image that you have upsized and sharpend?

Would be nice to see this actually.

apefos

I already uploaded an image in the post number 7 showing the difference between "Instant HD upscaling" and "default upscaling"

Both images was identically sharpened in ACR before the upscaling. The settings for tweaking these images are described in my posts here in this topic, read them all.

N/A

7D. 600D. Rokinon 35 cine. Sigma 30 1.4
Audio and video recording/production, Random Photography
Want to help with the latest development but don't know how to compile?

Luiz Roberto dos Santos

Please 1% and Apefos, see this document in which I compared several algorithms. The ACR and AE use the Bicubic algorithm, technically inferior to all others (except bilinear). The plugin red giant probably uses the Lanczos, Hungarian algorithm.
The best interpolation algorithm until now demonstrably is Smooth Spline. See it all in the document (see at 400%):

https://www.box.com/s/q1i471h34lqm50c1v930

Audionut

Quote from: apefos on June 07, 2013, 11:35:24 PM
I already uploaded an image in the post number 7 showing the difference between "Instant HD upscaling" and "default upscaling"

Yes, but I would like an untouched DNG so that I can perform my own tests please.
I own a 5D3 that doesn't have aliasing problems.  I guess I could run my tests on my own footage shot at a low resolution, but I envision that someone will come along and tell me that my tests don't count cause it was with a 5D3.

Roman

Quote from: jchristman on June 07, 2013, 08:40:32 PM
I just came across this optical filter solution for moire. I'm intrigued:

http://recordingsofnature.wordpress.com/2012/12/04/cyl-filter-moire-and-aliasing-filter-for-hd-dslr-video/

Since there are lines skipped causing the aliasing, the problem is that some light reaches one pixel*, some light reaches the gap in the middle, then some light reaches the next pixel across.
So what an aliasing filter does is blur the light coming to the pixels, so the lost colour information that otherwise gets sent to the 'gap' is included.

Which obviously reduces sharpness and contrast to some extent, but the idea of the aliasing filter is to reduce it by the exact minimum amount required for sensor size and how much line skipping it's doing so it's not noticably blurry...

Could you not acheive exactly the same thing, by shifting the lense focus by a tiny amount forward or backward from a 'perfect' focus and then sharpening the image afterwards?
I know that sounds pretty blasphemous to a photography crowd to intentionally misfocus, but maybe worth experimenting with.
Or just smear some butter on the front of your lense, hahaha.

* = I realise this isnt the correct terminology, not sure what the RGBG cluster type thing is called that recieves light on the sensor.

3pointedit

Infact this is exactly how I shoot computer monitors at work. Often you find a moire effect when the screen pitch matches that of the camera resolution (a form of extreme brick wall). The easiest way to solve this is to back off focus just a little bit. But sure you get a little contamination in the blacks/edges.
550D on ML-roids

Yoshiyuki Blade

Quote from: Audionut on June 07, 2013, 05:40:37 PM
I had good results with this back in the day.

If you're handy with avisynth, you can try changing the resizing algorithms in the script for better results.

Apply a good sharpener at the end of your script to break the laws of physics :P

Be prepared to significantly increase your processing time.  But hey, quality comes at a cost ;)

I think I fooled around with some AVS AA filters for this kind of video, but I don't recall getting anything satisfactory at the end. I probably didn't use them correctly though. I wonder if an implementation of Nvidia's FXAA could do anything about it. It's a post-AA so it doesn't require real-time rendering to work. I searched for info about it a while ago on Doom9's forum with not much luck.