How to remove a lot of film grain from DVD source

[DJ Alik]

Hi I am trying to get rid of film grain from the Devil's Advocate DVD. It's very grainy!!! Here are some source frames

http://www-users.itlabs.umn.edu/~rusi0006/frame1.tif

http://www-users.itlabs.umn.edu/~rusi0006/frame2.tif

http://www-users.itlabs.umn.edu/~rusi0006/frame3.tif

if the link doesn't load right away try hitting refresh.

I have already tried many different things. First of all undot(). That didn't do much. Then I tried various combinations of

Temporalsoften(4,8,8,mode=2,scenechange=10)
Convolution3d("movielq")
FluxSmooth(10,15)

Can anyone suggest me a filter or a set of filters to use on this specific problem. I would be very very grateful.

[digitalman]

I would be interested also in some information on this. I tried backing up a 70's movie transferred to DVD(supposedly digitally remastered) and was unsuccesful in cleaning up the grainy film.

[sh0dan]

Filters that might help you (as I don't have anything from the source it's hard to tell).

VagueDenoiser. This one can be very powerful at grainremoval in my experience.

MipSmooth - for softer grain removal.

C3D - also for soft grain removal.

Peachsmoother.

Temporalsoften - in general not to be used with high thresholds, as it mostly causes smearing on thresholds > 5.

[Boulder]

I've successfully used UnDot followed by FluxSmooth at the default values with some Hitchcock collection movies (The Birds for example). They keep the details very well.

You've also got a nice halo effect around the edges of the character:scared: A slight UnFilter, possibly UnFilter(-5,-5) would help there?

[DJ Alik]

Thanx for all the replies. But I am not a scripting guru. Could someone tell me what exact parameters to try in a script form. Thanx

[slk001]

Try using the CondTemporalMedian filter. You can get it here:

http://kevin.atkinson.dhs.org/temporal_median/

I've used it to clean up older films.

[DJ Alik]

I did what Boulder suggested:

Undot()
FluxSmooth()
UnFilter(-5,-5)

I think the result is pretty good. I didn't actually convert the movie yet but the output frame in VDubMod looks pretty good

here is the original frame:



http://www-users.itlabs.umn.edu/~rusi0006/frame3.tif

and here is the edited

http://www-users.itlabs.umn.edu/~rusi0006/frame3mod.tif

Does anyone think there is something more to be done to improve it?

Big thanx to everyone who replied.

[DJ Alik]

Ok. Not since I had a few minutes of free time I decided to test the filters on a small clip. See for yourself

http://www-users.itlabs.umn.edu/~rusi0006/sample.avi

I am not happy with results at all. Look at the background. Especially the walls and such. Lots of artifacts. I need some help with this. What should I do.

[Inc]

Boulder gave me a very good hint to Peachsmoother. So I build a combination using also TempSoften and a sharpener if needed.

so try.

- on noisy DVDs....

Temporalsoften(1,3,5,3,2).Peachsmoother()

- on noisy DVDs including sharpen a bit (not in your case!)

Temporalsoften(1,3,5,3,2).Peachsmoother().unfilter(60,0)

- on less noisy DVDs....

Peachsmoother()

- on less noisy DVDs including sharpen a bit (not in your case!)

Peachsmoother().unfilter(60,0)


Just try what works for you ...

Edit: Maybe you have to convert to YUY2 before using Peachsmoother by using
ConvertToYUY2()

[DJ Alik]

It looks a bit better but still there are some parts that are just horrible. Especially when the characthers are standing in front of a window and you can see the artifacts jumping around in the white background.

[morsa]

if you're talking about the macroblocks try using blockbuster or addgrain or something similar.

[Teegedeck]

In case you are encoding to XviD, from my experience it looks more pleasant to leave grain be. Try using the H.263 quantizer with Trellis quantization and quarterpel - it smoothes but still looks crisp. Nothing for 1-CD encodes though, as the noise of course decreases compressiblity.

[MrTibs]

For fixing the white halo problem, I would suggest the FixVHSOversharp filter. It will eliminate them completely. For DV source you will probably have to run it several times but the results should be much better.

[gizmau]

did you try the cpu-options of mpeg2dec3 allready?

[Fizick]

My source is not DVD (but from old 8mm film).
But how remove a lot of middle-size grain?
I try Vaguedenoiser with ugly results with high thresholds.
Any filter combination or algorithm?
Seems, the best result i have got with NeatImage photo-software.
Who know its algorithm ?
The speed is not important to me.

Sample image:
http://bag.hotmail.ru/images/grain.jpg

[morsa]

I guess this is a task for Frequency Domain filter, and I guess you are the right person for this, Fizick:)
Neat image uses wavelets AFAIK, but don't know its algo.
Also Restoreinpaint has a big library of filters to correct many types of problems.
Doesn't bilateral filter from Trevlac help?

Here are some links :

http://www.j2000.org/

http://www.ece.uvic.ca/~mdadams/jasper/#software

EZW codec.Uses Wavelets.

http://pesona.mmu.edu.my/~msng/EZW.html

Original Matlab:

http://perso.wanadoo.fr/polyvalens/clemens/ezw/ezw.html

Wavelet image constructor Kit

http://www.geoffdavis.net/dartmouth/wavelet/wavelet.html

Hope this helps..

[d'Oursse]

Fizick: you could use golddust to remove such noise. It's a quite impressive plugin. And it's slow :D

[Fizick]

The Golddust is great, but not sufficiently slow for me. :)

Morsa,
I already tryed some algo from Restorepaint (thanks for your point to it some time ago) , and writed some quite working script for Black Top Hat algo simulation. I use Masktools, but it is lacked in true erode-dilate for not 1-point objects.

Also I try to tweak Vaguedenoiser, but without great success.
About wavelets, i found a link:
http://taco.poly.edu/WaveletSoftware/

[morsa]

I name it again, just because it has many new functions, like AOS Barash denoise/deconvolution and I guess that that coupled with some Susan denoising or even a bilateral would give good results for your source.
Thanks for the link:)

[Didée]

This is that kind of threads making me believe I live behind the moon.

Go on, guys, go on :)

- Didée

[kassandro]

As Boulder mentioned Undot, I might take the opportunity to preanounce my filter RemoveGrain. With RemoveGrain(threshold=1) it behaves exactly as Undot (but a lot slower). With RemoveGrain(threshold=2) it still doesn't cause any artifacts, but does a lot more denoising. With RemoveGrain(threshold=3) thin lines may be destroyed, but usually the output remains fine even with the maximal threshold=4. Thus RemoveGrain is for fine grain, not for very crude dirt like in Fizick's picture (I couldn't download DJ Alik's pictures!). It was designed as a precleaner for RemoveDirt (http://www.RemoveDirt.de.tf), whose motion detection may be confused by grain. While the filter is quite stable now, documentation is missing and it is too slow. Right now it makes only 21 fps with a 720x576 clip and threshold=2 on my 1.3 GHZ Celeron and with SSE I hope to raise the speed beyond 100 fps. I hope to release the plugin until the end of July.

[Fizick]

Kassandro!
ONLY 21 fps? :scared:
I want write a similar filter with about 1 fps! :)
See my script impaint:

AVIsource("h:\smena0.avi")
loadplugin("masktools.dll")
function impaint(clip s, int difwhite, int difblack) {
opening=s.inpand().inpand().expand().expand()
maskw=YV12Subtract(s,opening).binarize(127+difwhite)
deflated=s.deflate().deflate().deflate()
closing=s.expand().expand().inpand().inpand()
maskb=YV12Subtract(closing,s).binarize(127+difblack)
inflated=s.inflate().inflate().inflate()
impaintw=maskedmerge(deflated,s,maskw)
impaintb=maskedmerge(inflated,impaintw,maskb)
return impaintb
}
converttoYV12()
dw=30
db=30
impaint(dw, db)


I want try your prelease (PM or e-mail)!

[kassandro]

Fizick, you can download the plugin here (http://home.pages.at/kassandro/RemoveGrain/RemoveGrain.dll). There is essentially one variable called threshold with values between 0 and 4. 0 = do nothing, 1 = slow Undot, 2 = default mode. The higher the threshold the more denoising and the slower the filter. The speed also depends on the amount of noise. The algorithm is a substantial refinement of that of Undot and will be described in detail in the documentation. With debug=true the number of pixel changes is printed to debugview. There may be up to three pixel changes per pixel (for Y, U, V values)(the debug output will be changed in the official release due for end of Juli). Please don't distribute this plugin any further.

The link seems to work only, if it is typed directly into the internet explorer.

[Fizick]

Thanks, I download it.
Will try tomorrow.

[Fizick]

Seems, maximum threshold=4 is too low for my clip. :(
By the way, i decide to release a tweaked version of Vaguedenoiser.
(the speed is surprisingly high, about 3fps :) )
But grain is still on frames. :(