New fast denoisers: DCTFun4 / DCTFun5 [Archive]

View Full Version : New fast denoisers: DCTFun4 / DCTFun5

Soulhunter

9th August 2006, 12:45

As the title says: Two new fast denoisers, DCTFun4 and DCTFun5, both do hard thresholding of a complete 4x4 [2x2 for DCTFun5] ict transform... DCTFun4 is a bit slower than DCTFun5 but should give better quality, DCTFun5 is a bit faster but less powerful in noise suppression [they are simply scaled for different speed/quality goals]. Both should be able to beat most [all?] other speedy denoisers out there, visually and metric wise... But still, please post feedback regarding their denoising capability's, and maybe some speed comparisons as well!

- DCTFun4 (http://soulhunter.chronocrossdev.com/data/dctfun4b.zip)
- DCTFun5 (http://soulhunter.chronocrossdev.com/data/dctfun5_r1.zip)

The param description is in the zip...

Ohw, and again, plz dont ask for the source, Im not the author!

Tia n Bye

manolito

9th August 2006, 13:06

Thanks Soulhunter!

There is a typo in the link for DVTFun4 (please delete the "p" in "chronopcrossdev").

Cheers
manolito

Soulhunter

9th August 2006, 13:14

Thanks Soulhunter!

There is a typo in the link for DVTFun4 (please delete the "p" in "chronopcrossdev").

Cheers
manolito

Ohw, thanks for the hint! :]

Bye

IanB

9th August 2006, 14:20

Ohw, and again, plz dont ask for the source, Im not the author!Any chance of a hint as to who the author is. :)

Soulhunter

9th August 2006, 14:24

Any chance of a hint as to who the author is. :)
Same as for FRFun, Prunedtree! ;]

Bye

Mr. Brown

9th August 2006, 14:24

thanks for a new toy !

Boulder

9th August 2006, 14:28

A small bug: when I set the thresholds and have YUY2 material, the video gets garbled and Avisynth crashes with an access violation. This doesn't happen if I just use DCTFun4B() or convert to YV12 before the filter.

MfA

9th August 2006, 16:12

guldukat

9th August 2006, 17:58

The anonymous&binary thing is one thing, but could you ask him to mention the originator of the algorithm if it isn't completely his own idea? (I hazard a guess this is based on Onur G. Guleryuz's work.)

1. yes, any decent denoiser will share resemblance with onur's works
2. AFAIK onur's work involves optimised weighting schemes for overcomplete representations that this filter doesn't use (so it's closer to donoho's work... but not on wavelets)
3. the only value in this filter is the actual implementation, tweaking and hand assembly

ps: why the hell are you not irc anymore MfA?

MfA

9th August 2006, 19:31

Onur's method can work with overcomplete transforms.

He also described a rather simpler version using a sliding block DCT though. Basically he just weights results from overlapping thresholded transform results based on the number of significant coefficients, the less significant coefficients the greater the weight. Which was a rather inspired idea.

Dunno why Im not on irc anymore, probably because I haven't really coded anything usefull lately :/

guldukat

9th August 2006, 20:40

foxyshadis

9th August 2006, 21:14

It's nice to have a faster competitor with fft3d, but it'd be nicer if some of the minor bugs were also fixed in the frfun family. >.>

Mug Funky

10th August 2006, 02:50

hmm, the results look similar to a median filter if the thresholds are pushed up...

but i'll give this a go. i'm sure it helps compression quite a bit.

Didée

10th August 2006, 05:49

Comparing with fft3d isn't really appropriate. It's a nice fast spatial filter, not more, not less. fft3d is a little bit more, iirc.

And indeed, results remind a bit on median filtering. In fact, it seems to me results are somewhat similar to SPresso() (http://forum.doom9.org/showthread.php?p=743559#post743559) ... only that the latter runs 3 times faster as dctfun4b ... :whistles inwardly:

acrespo

10th August 2006, 13:29

I use Removedirt in combination with dctfun4b(8,7) in anime content and the results are very good. The noise has been removed and the details preserved.

Soulhunter

12th August 2006, 17:13

And indeed, results remind a bit on median filtering. In fact, it seems to me results are somewhat similar to SPresso() (http://forum.doom9.org/showthread.php?p=743559#post743559) ... only that the latter runs 3 times faster as dctfun4b ... :whistles inwardly:
And that from someone who's famous for 1fps functions... ;]

But are you sure that the results of SPresso are really comparable?

Bye

Didée

12th August 2006, 19:38

But are you sure that the results of SPresso are really comparable?
To some extend, yes. At least from a few tests I did, which were quite few indeed ... because you know that

- I refuse to use filters without an explanation of "mode of operation".

- We're an open community here. I'm not for underground activities, keeping wannabe-elitist secrets, and every now'n'then trying to heighten by throwing some "try this" pittance ...

Keep in mind, Avisynth is one of the most incredible open source projects. (What do you think the charge would be if it were a commercial app ... ? ) :eek:
In this context, secretiveness just doesn't fit the game.

Soulhunter

12th August 2006, 21:02

We're an open community here, yeah! Thats exactly the reason I spend my time to pull this filters "out of the underground" by asking the author if I can host em on my site and open threads here to share em with the community... For the explanation of "mode of operation" thing: Err, I cant tell you stuff I dont know, heh? All stuff you need to know to use this filters can be found in the readme's and threads! If you want additional info, ask the author... I dont feel like bugging him for stuff I dont need/want to know! Lastly, keep in mind that the alternative would be that this filters stay "underground" and get only used by 3-5 weirdos like me... You dont like em, then dont use em, fine... but please stop the nagging, it starts to get disturbing! :]

Thanks n' Bye