Avisynth: Ads() a function using a masked denoising etc
 

----
---------- Splitted from Thread "MA Script for one Movie DVD" ---------
----

Well I see that Removegrain does affect details in a bad way, so I did let perform that filter on a masked area where the details are left OUT.
We also do often use Temporal denoisers on clean sources but these "could" do harm on details in hairs and skins, so I did let perform a temporal work also on that masked area.
And finally to get even the details a bit more sharpened I used on thats same (now inverted) mask a general sharpener.

Here's the function to test:

ADS - Adaptive Denoiser & Sharpener

EDIT:
Latest version and explanation can be found here:
http://www.kvcd.net/forum/viewtopic....r=asc&start=52

....

where To is the Top border and Bo the bottom Border to get rid of (you can detect (not set!!) them using DVD2AVIs clip/resize function).
Ov is the Overscan.
Lb means if set to "true" then a Letteroxing from Anamorphic material to 4:3 will be performed.

So ADS(72, 72, 2, true) deletes on the Top and The bottom the orig 72px and 72px borders and does apply an overscan of 16px where the right borders for matching the final output size will be added afterwards (imge will be centered).

To or Bo have to be MOD8! So each border to delete has to be devidable by 8, as removegrain() only processes mod8 sources.

As I do determine the filter to its parent-plugin, for instance "Masktools_Edgemask(....)" ... use latest version of Avisynth as that version only supports that.

Plugins Needed:

RemoveGrain()
www.removegrain.de.tf

RemoveDirt()
www.removedirt.de.tf

Masktools
http://jourdan.madism.org/~manao/MaskTools-v1.5.1.zip

Hi Inc. :),
Just tried ADS for the 1st time.
I downloaded and extracted Masktools v1.5.1, RemoveDirt v0.6 and RemoveGrain v0.6.1 to my avisynth plugins directory.
Then I used:
And I got a "Out of memory" from any tool I used to open that script.
You name it: MPC, WMP, VDMOD, etc.
I even erased all the stuff I had in the plugin dir and only left these 3 that I already mentioned and mpeg2dec3.dll.
Then I watched carefully your script and I decided to add unfilter.dll to the plugins dir.
But no way.
No way until I erased ADS.avsi which was copy pasted from this thread.
Can someone else replicate this?
Cheers

Try ro use that function in your script and not via avsi, maybe that helps :?

I used it yesterday and had no problems.

Are you shure you have all the latest builds (masktools avisynth!!). As some recent avisynth betas are known for these memory errors.

:wink:

EDIT:

I found this:
http://forum.doom9.org/showthread.php?s=&threadid=82435

Maybe you should try that masktools 1.49 or so, but I do hope that it gots the same commands/parameters.

http://jourdan.madism.org/~manao/?M=D

Maybe that one?
http://jourdan.madism.org/~manao/MaskTools.dll

or this
http://www.avisynth.org/warpenterpri...l_20040423.zip

@incredible
Thanks a lot! Just tried your ADS routine and it looks VERY good! :D
The only "problem" I have is that the output seems to be 704x576 instead of 720x576 as I usually use. Not a big issue, I know, but if it's fixable... It's just that I feel more comfortable with using 720x576. :wink:
Thanks again!

PS. Since the output isn't 720 it looks as if ADS is cutting about 16 pixels from each side of the frame and then zoom in on it?! How do I fix that?



@rds_correia
I added a "}" to the end of the script since VDub said it was missing :wink: . And now it works for me.

@ Audioslave

Change the line into ....

As the function in there detects the source width so it cuts in case of 720 width 8px on each side, ... as I do see no sense in re-encoding 720.

I just fixed the script according to the "}"

@incredible

Yes, you might be right :wink: .
However, this is not really the issue in my previous post...
When comparing a simple LanczosResize with the ADS script (both with overscan 2) ADS clearly shows less of the original frame than the LanczosResize. I hope you know what I mean :oops: .
If not, I'll post some screenshots when I get home from work :wink: .

Hello :),
My thanks to both Henrik and Andrej for fixing the function.
I'll let you both know how it goes when I get home.
Cheers

@incredible
I tried ADS on the Scandinavian (PAL) version of "The Passion Of The Christ". It looks like a whole new movie! :D Well, the plot is the same 8O :) but the picture is soooo much better with ADS! The original is VERY grainy and looks very "VHS" but after running the source through ADS it now looks like a DVD! :lol:
However, ADS sometime is a little too tough = loss in detail, for example beards and hair...

P.S. I still haven't figured out why the resizing is so wierd... :cry:

Two notes about this ACS function :

- Removedirt seems to give problems to CCE when parameters are used. Does ACS works correctly ?

- please check this thread I did on my last attempts with RemoveDirt
http://www.kvcd.net/forum/viewtopic.php?t=14141
(this behaviour is perhaps reduced with usage of thresholds).

Resizing is not wired at all! ;-)

a) Almost every DVD Source I do keep at Full PAL, just cropping from 720 to 704 as it makes sense. On long Movies like 3h ones I sometimes choose letterboxing (Anamorph :arrow: 4:3 incl. Borders) to gain from better compression.
b) When Overscanning I like to choose Overlayed Overscan as many movies do come in 2.35:1 which means when using resized overscan, the moviearea even would get smaller on a regular Tv. So Overlayed is the best choice in case of DVD Movie sources.
c) And for shure I do crop the orig black borders before so the filtering affects only the main movie area, means better detection for some type of filters and for shure more speed as the effective movie area in a border cropped state is cropped effectively!

You can easely enter the script and add your own mods:

Here you see the parameter determination and the variables out of the commands/syntax and the determining of the defaults IF no parameters are set at like just "ADS()".

Ov is Overscan taken as 0 or 1 or 2, but we want real pixels to use with the follwing cropper, so we just multiply it by 8, so in case of Overscan =2 the result is 2*8 = 16px on each side .
We do parse the sources ORIG width and Height and set them to W and H.

The first line checks if W (orig width) is 720, if yes an extra cropping of 8 at each side is determined as cr=8, if not then cr=0. Followed by the cropping job using Crop(...)


Here we do set the edgemask and finally we do blur it so little details wont begin shimmring cause of to hard edge masked.
(The nf=Ydiff..... line Ill delet as its useless ;-) )


f1 is the first filter job. It includes The spatial softening job Removedirt and Removegrain. Removedirt wont harm edges now as it will ONLY be performed on plain areas - cause of the edgemask ;-)


The Second filter Job is f2 where Unfilter does enhance the edges in the movie



Here we do apply the filtered areas to the original image depending on the mask, while the second merge is inverse last,f2,m instead of f1,last,m


Here I do check if the Lb boolean variable out of the command ADS() is set to true, if yes a anamorph to non anamorph resizing will be performed, if not, then leave the image as it is.


Finally I do check how much borders are needed to fit the final active movie area "centered" to the ORIGINAL size of the beginning inlc. respecting the cropping from 720 to 704.
So finally the borders will be added.

END.


You can for instance

"f1=removegrain(8).RemoveDirt(mthreshold=125,cthre shold=30,pthreshold=20)"

To

"f1=SpatialFilter(.......).TemporalFilter(.......) "

And

"f2=unfilter(0,80)"

To

"f2=SharpeningFilter()"

But be shure you use YV12 supporting filters, if not you need to perform colorspace changes which do let drop the color quality and for shure more cpu power is needed ;-)

@ Phil
I do see your last post now in here.
And for shure the pics in the link you pointed me to 8O (seems I lost that new "good bye" thread. Me and boulder (IMHO) haven't seen such problems, but maybe we didnt recognise them well !?
So do your testings with the thresholds I used, if they dont work, do choose other thresholds.
If that still wont work, we trash that one finally and change to a different PURE temporal denoiser like Tempsoften, ATC, TempCleaner etc.
The place where to change you can easely see above ;-)

Also in case of the Spatial denoiser/softener, maybe just Deen("a2d",x,x,x) "could" work ... BUT be careful as deen() in a conditional scripting will crash! in my cases! the result is a "access volation ....." as this filter deeply isnt correct coded IMHO. But there are also other pure spatial softeners around, but yep, Removegrain() does its Job very good. So testing temporal filters should be a real job.

@incredible

:lol:
Thanks!

You're the man... 8)

As I use CCE for my KDVD and RemoveDirt gives problems to ECLCEE, that is a "no way" :-(. I will try with tmpgenc just for testing purpose.

But for me RemoveDirt suffer from a big problem : there are TOO MUCH parameters. And I'm not sure it will be easy to find values that will face to all situations.

Yep I forgot that! So we should change to another temporal filter in case of EclCCE.

Well I do think we both (and others to) got a testing spirit, so thats why you also found out optimal settings for convolution3d() ;-)

Whats about a temporal used StmedianFIlter() or even Convo3d() at temporal settings ?
I havent test them for a long time 8O 8)

I have never had any problems with RemoveDirt in CCE or ECLCCE...

Try to use "RemoveDirt(mthreshold=50)" insteed of just "RemoveDirt()".
EclCCE answers : not abble to load the script. (CCE 2.67)

Phil, I do assume you did NOT your last screenshots using my masking function???

*IDEA*

Because this function does avoid that details/Edges even will be touched by the spatial or temporal smoother/denoiser! So that Arc in your test screenie wouldnt suffer (just IMHO).
So do use Removedirt() in that function without any parameters and do a test, as Im now at work (you probably too?!)

No, I wasn't connected for all the we so I couldn't do that :-)
(note: I give the script in my post, there is nothing more than the 2 filters).

I am but all the sources are on my laptop :-) So I can try that.

Yep! You quoted that :lol:
Nice company youre working in .... do they need another employee ???
8) [/quote]

Ok, just did it there :
http://www.kvcd.net/forum/viewtopic.php?p=103824#103824
Note : I used ECLCEE with no problem even with the parameters used in RemoveDirt.

I do not get that error message. I'm using EclCCE 1.81 with CCE SP 2.67.00.27. Whenever I've had problems with with plugins I always uninstall avisynth and install the latest version(v2.5.5 final just came out btw). My project files were created using DGIndex 1.0.12 by neuron2. Here is the script I used:

This is the only diff between your situation and mine but I don't think this can be the guilty. I will try to check.

Hi guys!
I just made a test with FluxSmooth (the new version) and UnDot instead of the Remove-filters, and I really liked what I saw! :wink:
This is what I used:

And this is my script:
I only wanted to inform you...


BTW Is there a way of using, not overlay overscan, but the other kind where you shrink the size of the frame? With ADS that is...

Try setting Blur(0.2) to Blur(1.5) and unfilter(0,80) to unfilter(80,80) .... Im testing right now and it seems that when doing this the details do even more come out .

Also I did find out that removedirt at default gots HEAVY settings according to motionthreshold! Do lower it to approx 50 if using it:

f1=removegrain(8).RemoveDirt(mthreshold=50,cthreshold=30,pthreshold=20)

I changed the "roberts" Edgemask architecture to "sobel" as it detects better the edges in a way I want.

With the rise of blurring the mask using (1.5) and rising the sharpener to 80,80 I got nice results!

Two Commands are compared


The pics:

http://www.digitalfaq.com/archives/error.gif

http://www.digitalfaq.com/archives/error.gif

http://www.digitalfaq.com/archives/error.gif

http://www.digitalfaq.com/archives/error.gif

http://www.digitalfaq.com/archives/error.gif

wow incredible, that ads filter is quite something! I gave it a small test and I have to say that I like what I see. I do think I'm getting larger file sizes with it, and encoding takes much longer, but the results are impressive. I have two comments right now:

1) you mention to make sure the top and bottom borders (To and Bo) are mod8 because RemoveGrain() requires that. But I just finished encoding 22 episdoes of Alias where I removed 20 pixels from the top and bottom, and used RemoveGrain(mode=2) without any problems!

2) I have only one problem with RemoveDirt, and I've mentioned it many times. It can cause "twinkling" of small objects such as stars, background lights, etc. during camera pan scenes. To see what I mean, encode a scene where you get a pan of a night sky, or a pan of a skyscraper at night. With removedirt(edited) you WILL see this twinkling, but without it you won't. I really wish someone else could verify this for me cause I think I'm going crazy! 8O

Hmm, another thing I noticed about the Remove-filters -> During panning they (or one of them) creates stuttering in the movie. In other words, the panning is far from smooth. Seems like one (or more) of those filters isn't adapted to motion... :?:

@ J-Wo & Audioslave

In the latest version of ADS() I used ...

RemoveDirt(mthreshold=50,cthreshold=30,pthreshold=20)

As the motionthreshold at higher settings than 50 does "just" kill details and some kind of a bad looking motionblur happens! 8O So the latest setting maybe solved that? Test are welcome.

At your point 2) you mention the problem of "twinkling". But one time related to RemoveDirt and at the end the other time related to Removegrain :wink: :arrow: :?: Could it be you mixed up both filter names (like I did in the past) ? :)

Related to filesizes:
Yep, as said, this funktion gots its focus primary on detailenhance and quality. And as we gain from a luxury MB space on a DVD-R Media this doesent matter in case of DVDrebuilder, you agree? :wink:
But as also said, you can easely lower the unfilter(80,80) to some lower values like (30,30).

J-Wo, a cropping of 20 & 20 will result in a total cropping of 40 which is devidable by 8, thats why it works, so it was my mistake to say "cropping has to be mod8" as it should be "the final cropped movie area should be mod8". So as you live in NTSC land you got a height of 480. So 480-20-20 = 440 and thats MOD8 and thats the area processed by RemoveX() :)

Do a test by cropping for instance 26 and 26 and the error will occur.

Maybe this filter is something to test?
http://forum.doom9.org/showthread.php?s=&threadid=80834

haha sorry for the confusion, I corrected my above post. I did always mean RemoveDirt, as I remember it is the 2nd filter in the combo RemoveGrain.removedirt.

I did do an encode with your newest settings of mthreshold=50. You are right that edges no longer move but I did see a brief scene in which some slight twinkling was noticed. I do believe it was less pronounced than removedirt at default settings. I compared ADS to my simple script of removegrain(mode=2).temporalSoften(2,4,6,6,2). My source was anime, but I'll try and find one that is live action to post.

Since I see you are testing with Kill Bill, maybe you could try the end of the movie which is a nighttime snow scene. I believe there are some shots which look up into the sky and pan the stars. Please test to see if you see the same problem as me! :roll:

Ahhhh thanks for that, makes a lot of sense. Is that ALWAYS the case when people say filter x needs mod8 height/width, such as BlindPP?

Be carrefull : this is what I used to use month ago and I dropped fluxsmooth because it is too heavy on the smoothing (it is at the origin of a lot of DCT blocks in the plain areas).

For sure the result are good. But can you tell us a little about file size ? Because in my only test (for the moment) of your function, the file was 18% bigger 8O

I did an encoding test on that KillBill1 last night and got approx 5-7% bigger filesize :? Encoded in 704width anamorph and I added 16px overscan to both sides and kept the orig size of the top/bottom borders (added afterwards again in that script).

More details do need more bits, so someone has to do his own compromise.

But shure, every movie is different. :wink:

Well we'll see, .... also it depends on HOW long is the movie and how much bits in avg are needed. Cause when doing DVDrebuilder workouts with the usage of ShrinkDVD (menue shrinkage) and Vobblanker (geting rid of Trailers) before I get that low Q values in CCE that this little rise IMHO you wont notice, like a jump from Q10 to Q 15. But this is just the beginning of the tests with this script. It was just the beginning of the usage of masktools, .... and I got some other Ideas totally different than this script. :)

To clarify the things, I was speaking of myt test on your very first version of the fucntion (the "old" one, that have 2 days now ;-)). I did not try yet the last enhancement.

Why didnt someone advise me that a conditional/scriptclip environment isnt needed in my function :banghead: :banghead: :banghead: :screwy: :lol:

Not needed that even more power consumpting scriptclip routine.

So I rewrote the function.

AND! A new Idea :!: :!:

Why do we denoise an area at full resolution IF there are (cause its masked!) NO necessary details ???? :arrow: No reason.
So I pulled out my old idea of ONE-HALF-ONE, where I recognised that denoising is MUCH more efficient if the filter gots LESS pixels to handle.
http://www.kvcd.net/forum/viewtopic.php?t=6863

So AS I did mask that area to treat by a denoiser I easely can reduce its resolution to half, denoise it better and bring it back to its origin resolution WHERE I use bicubicresize using the parameters 1,0, which means an extra spatial softening, so no extra spatial softener is needed!!
Almost every important detail will be kept as the details have been masked before and will be added back to the image finally! :D

Second... IF reducing the to-denoise-resolution, I choose reducing the HEIGHT by half! Why? Because IF I do denoise temporal a noisy area at half WIDTH resolution and scale it UP again, I would suffer from enlarged temporal denoise artefacts when the camera moves sidewards in a scene and if a camera moves its the case that almost 95% of all movie pannings are done horizontally. So keeping the horizontal resolution when doing temporal jobs would be inevitable.
ALSO .... most Details seen by our eyes are vertically as we do see horizontal sharpeness more sensitive: Best Example Hairs! Mostly, hair does got horizontal details (hair falls down) and it doesnt suffer from blur if we do interpolate vertical as the details of hairs are separated by our eyes in their width. ;-)


So heres the new function and tests are welcome.

EDIT:
Latest version and explanation can be found here:
http://www.kvcd.net/forum/viewtopic....r=asc&start=52

....

The syntax is the same:

ADS (clip c, int "To", int "Bo", Int "Ov", Bool "Lb", int "sh", int "th")

You see the parameter "sh" is added which stand for sharpening, means "0" does not apply any sharpening/detail enhance and 10 is the most (you shouldnt go above 8 ).
And the detail threshold "th", that one stands for the mask/detail threshold

So ...

ADS(72,72,2,true,6,8 )

Means ..

- do pre-crop the original black borders at Top and Bottom by 72px each,
- and do apply and overscan of 2 (=16px on each side)
- true means that a Letterboxing from anamorph 16:9 to 4:3 will be performed
- 6 stands for the sharpening factor (max 10)
- the 8 stands for the detail threshold, means the more big grain is present, the bigger thresh should be set. On DVD clean materail like 4 and on trashy old VHS captures maybe about 12


I also tested it on a shitty old VHS capture using a "th" of 12 and it did wonders. On almost clean DVDs 4 - 6 is enough.

And its quite fast as temporalsoften is performing on the half of pixel information (well the speedgain could suffer on the other hand cause of the masking job)

Damn it Inc. :D
You know I can't script otherwise I would have told you :mrgreen:
Great findings buddy.
Keep those nice ideas coming ;-)
Cheers

holy crap inc! You are a scripting wonder. Gonna give it a try today, along with DVD-RB which I have yet to give a shot. Do you think this script is too complex to be used with that program? I thought I read somewhere that it doesn't like removegrain/removedirt...

incredible,
I tried your new ADS function and it seems to work well. Only a quick test.

@ Rui

:lol:

@ J-Wo

As you can see in the function the removedirt is NO longer present as its now temporalsoften which does the denoising and the parameters 1,0 in bicubicresize which do the spatial softening.

So the filter which are NON avisynth native are masktools, unfilter and undot.
I do think there will be no Problem with Rebuilder (well in my cases not) :wink:

@ Peter

What hasAnamorphic to do with "processing pixels in a frame"? ;-) No Problem as I did tests on anamorphic "KillBill1" and non anamorphic captures.
If you want to keep it anamorphic just set the "lb" boolean to false.
The pre-cropping is up to you, no matter if its anamorphic or not. :)

incredible,
I already edited comments out. I did not read your instructions carefully enough.

incredible,
I did a few more tests on ADS and noticed that an anamorphic 16:9 720x480 input is changed to a 704x480 output by ADS. Since this function is designed to get the best possible image quailty, I would think that it would be best to keep all of the image information that is available. Perhaps you could add an option to not reduce 720x480 image width.