Fractal denoising: FRFun

[Soulhunter]

Ok, here (http://forum.doom9.org/showthread.php?t=109925) is the original thread and here (http://soulhunter.chronocrossdev.com/data/frfun3b_rev3.zip) is the binary... Please post feedback regarding its denoising capability's [comparisons to other denoisers, situations where it works well, situations where it fails...]

Oh, and dont ask for the source, Im not the author of this filter... ;P

EDIT: Look here... (http://forum.doom9.org/showthread.php?p=825604#post825604) and here (http://forum.doom9.org/showthread.php?p=862358#post862358)!


Tia n' Bye

[BigDid]

Hi Soulhunter,

Thanks for the good work. FRFun seems great and quicker than FFT3dfilter :)

I am a total noob concerning avisynth filters, could you elaborate more on FRFun specially concerning the parameters ?

I have seen interesting things on your page, could you consider adding a few denoisers in your denoiser comparison? I'm thinking about Fft3dgpu (to see speed advantages and if PSNR is equal or close to Fft3dfilter) and Removegrain, maybe mode7 or 8, to see if close or quicker than degrainmedian ?

Thank you for your time.

Did

[Soulhunter]

FRFun3b(T,Tuv,S)

The first parameter "T" is the luma threshold the second "Tuv" the chroma threshold (both are floats). As most sources have less chroma noise than luma noise its a good idea to keep the chroma threshold lower than the luma threshold! The last parameter "S" is the subsampling: More subsampling = less processing = faster, but also less precise! Its a "speed -vs- quality" trade...

For the denoiser comparison: Well, Im still looking for a better metric to measure the denoising quality! MSUs blur metric (http://forum.doom9.org/showthread.php?t=107214) looked promising, but after some tests I discovered that its useless (at least for my purpose). But as soon I find something better, Ill start a bigger and better comparison... So, stay tuned!


Bye

[Alain2]

Oh so it's float again ; is speed as good as previous 3b version with int parameters ? (I know, I will test, but a small indication from you is always nice :) )

What is the address of your webpage as mentionned by BigDid ? I only see the irc link in your signature...

[Soulhunter]

Oh so it's float again ; is speed as good as previous 3b version with int parameters ?
Yes, should be ~ the same!

What is the address of your webpage as mentionned by BigDid ?
Hint: The link to the frfun binary points to it!


Bye

[zlab]

Hi,
Any suggestion for lo, medium noise filtering parameters? pls.

[Arlong]

That's a nice filter. I tried it with FrFun(3,2,2) and this is what I get:

Source
http://img161.imageshack.us/img161/6600/vob5sk.th.png (http://img161.imageshack.us/my.php?image=vob5sk.png)

FrFun
http://img507.imageshack.us/img507/4245/xvid8xr.th.png (http://img507.imageshack.us/my.php?image=xvid8xr.png)

It removes well grain and kills noise without harming too much detail. Well, maybe hair looks a little overfiltered.

FFT3dFilter(sigma=10,sigma2=0.9,sigma3=0.9,sigma4=6) is worse in terms of noise reduction: the walls are still crowded with grain. Detail level seems to be almost equal in both pictures.

http://img162.imageshack.us/img162/7460/withfft8hf.th.png (http://img162.imageshack.us/my.php?image=withfft8hf.png)

[Disabled]

For me it looks like Frfun did a much better job. It removed more noise and kept more detail at least her cloth look more detailed than with fft3d.
I guess I will do some tests myself now!

[Mug Funky]

remember that fft3d can go temporal (in several ways too... bt=0 is my favourite but one has to be very careful with sigmas).

btw, i find that this is a good one for speed:
frfun3b(6,6,3)
degrainmedian()

i noticed a tendency toward "salt and pepper" jaggies when subsampling is increased, but following with degrainmedian will reduce those (this is more for the encoder's benefit than the viewer).

good filter - i can't wait to see the bollocks optimised out of it and a temporal mode introduced :). with motion-compensation it could be a real noise killer, but i'd only use it on really big frame sizes because it appears to eat details for lunch if abused.

[Serbianboss]

@Mug Funky

i am interesting when you use degrainmedian() command thas degrainmedian work with default parametars like this (limitY=4,limitUV=6,mode=1) or not?

What all of you suggest parametars for little noise and with much noise on interlaced film?

[AVIL]

@Soulhunter

Good filter. It cleans well with default parameters (rare thing). It loss any detail (e.g. shadows in faces) but generally is not noticeable. I think that a temporal behaviour could preserve some detail but I have not clue how to do it. It is very good fighting blocks from mov files. Tend to be slow but not in excess. Anyway is another tool for the box. Thanks.

[levi]

link not working?

[Soulhunter]

no, link working!


Bye

[guada 2]

" Soul"

It is a very good filter, but it is a little slow.
:thanks:

[Heini011]

current binary:

FrFun3b(3,2,2)

avisynth open failure:
script error: Invalid arguments to function "FrFrun3B"

[foxyshadis]

avisynth won't autocast int to float or vice versa. Use 3.0, 2.0, 2.0. >.> (I could have sworn this was working before, but perhaps not.)

[Soulhunter]

Humm, frfun3b(3,2,2) works perfectly here, I just downloaded and re-installed the dll from the link above to get sure... :\ Maybe its related to the avisynth version (I use 2.56)!? Oh, and btw, the last param (subsampling) is int, not float, so it should be frfun3b(3,2,2) or frfun3b(3.0,2.0,2), but not frfun3b(3.0,2.0,2.0)!

EDIT: Just noticed Heini011 wrote 'script error: Invalid arguments to function "FrFrun3B"'. But it should be "FrFun3B" not "FrFrun3B"! Leaves the question if it was a typo in the avs (problem solved) or the post (problem not solved) ^^


Bye

[AI]

for me frFun similar MSmooth, but easier (less parametrs) and faster
i.e. good planar denoiser for anime (becose anime have clear (sharp) borders, and frFun not spare blur (not excess clearance))
with temporal blur (http://forum.doom9.org/showthread.php?p=817752#post817752) for background I have very good denoiser for very noise anime.

PS Excuse me my bad English

[mgh]

it should be called a video cleaner-improves videos from commercial vcds in the sense that it reduces artefacts quite a bit and lets you see more details than before.
I have couple of problematic clips from vcds which i keep trying with various filters. This one takes the cake.
Thanx

[Zarxrax]

I just tried this filter, and all I can say is "WOW!"
I've tried it on some anime clips that have heavy amounts of film grain, and it can totally remove the grain! its amazing! No other smoother I have tried seems to even come close to this. This thing is amazing at preserving edges as well. The only problem I see is that it does seem to heavily blur some things that are real detail, if they arent defined very well to begin with, such as some texture on the ground, walls, etc.

[thoralf]

really nice filter, albeit slow as hell ... i prefer it over fft3d as it keeps more detail.
I'd like to know a bit more about its inner workings: how exactly do fractals help to eliminate noise and blocking?

thanks,
thoralf.

[BigDid]

...The only problem I see is that it does seem to heavily blur some things that are real detail, if they arent defined very well to begin with, ...
+1, even witout really strong params.

Could it help doing a soft/small sharpening before frfun, like a removegrain(mode=25) ?

I'll try to re-use the capture in which I discovered that or, if deleted, look at the next capture so not immediately.

Did

[AI]

Could it help doing a soft/small sharpening before frfun, like a removegrain(mode=25)?I use for prefilter for frfun3b, frfun3b too, with bigger parametrs...
for example:FrFun3b(4,3,6)
FrFun3b(3,2,4)

[Serbianboss]

Which parametars is good for noisy capures, and which parametars is good for low noisy captures?

best regard

[BigDid]

Which parametars is good for noisy capures, and which parametars is good for low noisy captures?...
Well,

Just re-use older posts from this thread as a basis:
-Default as a balanced start
-noisy captures, re-use AI params with 2 occurences, you have an example with pictures for:
FrFun3b(4,3,6)
FrFun3b(2.8,1.8,4) here (1st and 2nd pics): http://forum.doom9.org/showthread.php?p=821817#post821817
; lower params if not strong enough.
-low noise captures, higher params like FrFun3b(4,3,6) as a start?

Did

[Serbianboss]

thanks

[BigDid]

thanks
You are welcome.

As the use of this filter is new, it could be helpful for others readers for you to come back with the params you used with success :)

Did

[Soulhunter]

EDIT: Link removed, look here... (http://forum.doom9.org/showthread.php?p=825604#post825604)


Bye

[Alain2]

Would it be possible you make the help file just a bit more descriptive? What is lambda? The subsampling parameter has disapeared ?

[guada 2]

Hello Soulhunter,

Could you post the changelog please.
Just to see something..

Bye.

[Fizick]

Soulhunter,
please ask your second Ego :) for reference article.

[Backwoods]

*cough* mf *cough*

[tsp]

Would be nice to know if it was a candidate for a GPU version

[AS]

Let's just say the author would like remain anonymous and leave it at that...

[tsp]

http://www.f4.fhtw-berlin.de/~barthel/paper/spie_5266-5.pdf
maybe?

[Soulhunter]

Let's just say the author would like remain anonymous and leave it at that...
Yeah, and no source or reference article... for now, sorry! About the new parameter: Simply said, as higher you set the lambda param, as more aggressive smoothing you get... and yes, the subsampling param was removed! For the "changelog" thing: Well, the version jump from 3 to 7 indicates that a lot was changed... Lets just say it should be better than the old version, and the rest you have to find out on ya own!


Thanks n' Bye

[Soulhunter]

http://www.f4.fhtw-berlin.de/~barthel/paper/spie_5266-5.pdf
maybe?
Maybe, dunno! But it uses overlapping to make the subsampling less ugly iirc... >.>


Bye

[foxyshadis]

Hahaha, you guys treat it like it's illegal and you don't really want anyone using it, or if any description of the method gets out all the commercial companies will be rushing to implement his new algorithm and make millions off it.

Or maybe it's a code, and only the conspirators hold the key. ;)

I'm not annoyed, I just think the cloak & dagger around it is goofy.

[AI]

As for me, last version is faster;)

becose I use this script
FrFun3b(4,3,6)
FrFun3b(3,2,4)I found alternative on new version:FrFun7(1,4,3)
FrFun7(1,3,2)It's slower

PS. That I understand both version use alike algorithm cleaning
but different term finishing cleaning

PPS Excuse me my bad English

PPPS What is undocumented integer parameter "P"?
Then P > 0 new version is faster ;)

[AS]

Hahaha, you guys treat it like it's illegal and you don't really want anyone using it, or if any description of the method gets out all the commercial companies will be rushing to implement his new algorithm and make millions off it.

Or maybe it's a code, and only the conspirators hold the key. ;)

I'm not annoyed, I just think the cloak & dagger around it is goofy.

Look, if the author wants to release the code, then he will. Until then, it would be wrong not to respect his decision. So you can stop your conspiracy theories.

[Backwoods]

Someone is over defensive. Let's just enjoy the updates.

[Soulhunter]

Let's just enjoy the updates.
Exactly! :)


Bye

[Alain2]

What are the default values? Are the 3 param (Lambda, T, Tuv) floats? Ranges for these param are 0 -> +oo?
I appreciate the work and the share of it, but it would be nice if the readme could contain at least these info

[Soulhunter]

Ok, it seems there was a bug in FRFun rev1...

New version with better readme is on the way!


Bye

[Soulhunter]

EDIT: Link removed, look here... (http://forum.doom9.org/showthread.php?p=825604#post825604)


Bye

[Alain2]

Thanks for the improved readme :) (and the fixed dll of course ^^)

[Soulhunter]

EDIT: Link removed, look here... (http://forum.doom9.org/showthread.php?p=825604#post825604)


Bye

[Arlong]

Thanks SH. A very promising filter, indeed, growing better!

Just a little note: FRFun7 needs X and Y axis to be multiple of 8, otherwise it'll introduce green bars on the bottom and the right of the frame. So one has to place FRFun7 before crop or after resize (even between crop and resize, but cropping width and height with multples of 8).

Sorry for my english mistakes :p

[foxyshadis]

I tried it on some mildly noisy mpegs: Awesome. Best ringing ("mosquito noise") removal ever, with so little damage to detail. I haven't tried it on really detailed work yet though. Needs a sharpener at the end to really bring out the best, though, since it slightly blurs lines.

Only thing I noticed is that it sometimes jumbles up banding pretty good - in the original it was a smooth gradient of bands (meh), after it's all disturbed. It'd probably look great if it didn't have to get truncated back to yv12. Not always, just some I saw.

This needs to replace SPP deblocking; it's quite slow, but so is SPP, and this looks a lot better.

[AI]

WOW...
FrFun7 is "Spatio-temporal Smoothers", but FrFun3b is only "Spatial Smoothers" ;)

I know, becose need more memory for internal cache (I use SetMemoryMax)

i.e. may be good idea use "motion compensation" for better cleaning

PS and not filtred fist and last frames