DeFreq plugin - Interference frequency remover

[saibo]

A little update Defreq Version 0.7, April 28, 2006 - changed cleaning method: more frequencies are decreased near interference peak.
Chanded static linking of FFTW3.DLL library to dynamic loading.

Hi, I have a big problem. I tried your DeFreq Filter but it was not possible to eliminate the wavelike frequence erros. Is there any possibility to eliminate such errors?

http://img87.imageshack.us/img87/2633/testar0.png

Here an example:
http://rapidshare.de/files/26290105/Beispiel.m2v.html

Could it be a typical MOIREÉ effect?

[martin53]

Fizick,
would it be possible to expose just the forward / the inverse transformation as functions (like the left part of 'show')? That would make it possible to me to experiment with manipulations in the frequency domain - like overlaying a mask...

[Fizick]

martin53, sorry for late respond.
it is (almost) not possible right now.
'show' represents not raw transformation, but only real part of it, and logariphmed.
FFTW and I use complex float as transformation format.
1. You must store transformed image data somehow - well, it may be packed in any fictive format (e.g. RGB32).
2. You must provide means to manipulate this data - it is not possible in fictive format.
Try ask avisynth core developers to implement float (complex float) image format. ;)
It will not in avisynth 2.6. I already asked it. May be in myphological v3.0 :)

But It is possible to use integer storage format for DCT (not for FFT used in Defreq) with some scaling, but 8 bit is too anyway low IMO.
Try program C code to experiment.

[martin53]

Fizick,
a big thanks for your response!:thanks:
That's really bad news, but I kept feeling there was a problem like this.
Do you think it is possible that you multiply, or signed add respectively, a mask to the frequency domain? I am almost sure that it would be useful to subtract a small amount of the high frequency area - similar to what is done in subtle audio noise reduction in frequency domain.
And the multiplication with a mask could be an universal approach to your today maximum search and automated suppression of the peak in frequency domain, because one could just paint the suppression filter.

One more aspect: The transformation transforms something that is 'everywhere' in the 'time' or 'xy' domain to one point in the frequency domain, and vice versa. A blurred image due to camera shake contains the shake path (of the camera movement while the shutter was open) like an infinity of parallel paths in the whole image. In fact the image lacks high frequency along the axis or the path of the shake. And I wonder if this kind of information is also transformed to a certain location or localized figure in frequency domain that can be fixed. It should be like a dark area in the frequency domain. If so, FFT/CDT could provide an approach to sharpen images which suffer from this. Of course it is impossible to reconstruct all information fully - it was lost during the capture. But it might be possible to transform the ugly paths of the lit points into more standard shapes, and just leave some soft focus.

I know I can at least do some research in that direction with the help of 'show'. I'll add my results to this thread.

martin

[martin53]

I made an analysis script so everyone can examine the effects of simple blurring on spectral density for himself or herself.

pi = 3.1415926536
#====== USER INPUT =====

#Motion blur length (integer) [ 0 ... ]
# the value indicates the length ot the motion trail
# 0 produces an uncorrelated random noise clip
# variation changes the number of notch bands in the spectral density
blur= 7

#direction of blur in radians
# variation changes the angle of the notch bands in the spectral density
global phi= 0 #pi/20

#deltaPhi adds to the angle of the blur on its path and results in a bent motion
#values != 0 change the linear notch bands to elliptic(?) figures
global deltaPhi= 0 #1.0/blur * pi/4

#oversample factor (integer) [1 ... ] for the blur, compensates the Overlay(...) integer offset error.
#8 seems the best compromise for good correction and expense.
#no need for oversample with phi=0, phi=pi/2 etc
global oversample= 8

#clip size (integer) [32, 64, 128, 256 ... ]
#dimension of the quadratic original clip.
# 64 is fine to see the principles
dim = 64

#====== SUBFUNCTIONS =====

#mBlur performs the motion blur by recursively overlaying the semi-transparent, slightly shifted original
function mBlur(clip shifted, clip original, int i, int blurLength) {
Overlay(shifted, original, x=round(cos(phi+i*deltaPhi)*i*oversample), y=round(sin(phi+i*deltaPhi)*i*oversample), opacity=1.0/(i+1), pc_range=true)
return i >= blurLength ? last : mBlur(last, original, i+1, blurLength)
}

#====== MAIN SCRIPT =====

#create a random noise clip
BlankClip(width=dim+2*blur, height=dim+2*blur, color=$808080).AddGrain(2500)

#apply a motion blur to the clip - oversampled operation for better math precision -
oversample > 1 ? Lanczos4Resize(oversample*(dim+2*blur), oversample*(dim+2*blur)) : last
blur > 0 ? mBlur(last, last, 1, blur) : last
oversample > 1 ? Lanczos4Resize(dim+2*blur, dim+2*blur) : last

#cut off the border of the original clip where the motion blur could not be applied properly
crop(blur ,blur ,-blur ,-blur )

#make DeFreq() calculate the spectral density of the blurred noise
#and display the clip and its spectral density distribution
ConvertToYUY2()
stackhorizontal(last, crop( DeFreq(show=2), 0, 0, -dim/2, 0) )


My conclusions are
- movies with perfect quality have an even spectral density, i.e. a constant gray in "show" mode. (of course not one single picture, but the average over a complete movie)
- For real pictures with good quality, the limited bandwith makes the average of the spectral density dim to the right. Acoustically spoken, this would sound a bit dull. I verified that with some movies in place of the random noise.
- a remarkable motion blur has an affect that reminds my acoustical mind to interference with a slightly delayed echo, i.e. some frequencies are boosted because the original and the echo add with the same polarity, while the frequencies inbetween are extinguished because the echo has the inverse polarity and always adds to zero with the original.

martin

[martin53]

Fizick and Mohan,
I have some ambitions to enhance some of my photos and movies with spectral-based concepts.
I have photos in my archive where I could not keep the camera steady.
And when I shoot some video clips, it sometimes appears that during a pan, the frames are blurred because it is dark and the camera uses a long exposure time.

When I experimented with the script from my 1st follow-up, I was reminded of equalizers. While those equalizers (those with many sliders, or the parametrical ones which allow more precise control over fewer filters) only need to compensate a one-dimensional frequency response, the task is more complex in the two dimensional domain.

At least, beside the feature to attenuate some frequencies, all equalizers also need to amplify other frequency ranges.

:) So I kindly ask both of you: could DeFreq() and/or FFTQuiver() be developed further to allow amplification of frequency areas? As far as I understand, both are designed to kill some frequencies up to now.

The question needs to be answered: "Where to derive the filter settings from". I would like to present my thoughts here, it is not yet a perfect description for 1:1 implementation.

When working with photos, there is no chance to scroll through the frames and use DeFreq()'s show=2 averaging mode. Of course, the specific information in the image makes up a complex spectral density, and just normalizing to that must return random noise, not to a processed image.

But specific with the motion blur, there is one transformation which is neutral to the problem, i.e. could help: It should be possible to rotate the original image horizontally or vertically through the frame, and thus get a vast number of frames for analysis:

overlay(stackhorizontal(last, last), x=...)
# x = -1 ... -(width-1)

because the right and left edges usually don't tile nicely, the border should be darkened with some gradient. This is slow but it provides high quality which is needed for high resolution photos.

After that, a blurring should provide what I'd like to call the "defect mask" even from one source image. I'm still unsure if F2QMask could already use the inverse of this with some gamma change as "equalizer mask", because of the black-to-white range and the fact that F2Quiver multiplies (or convolves?) with the mask instead of dividing by (i.e. normalizing to) the mask. The proportional or inverse proportional characteristic of the multiplication is essential here!

When working with movies, any manual process and maybe even the slow approach above is impractical. But I think that it is not too complicated for a filter chain to generate the "equalizer mask" automatically on a frame-by-frame basis for DVD or HD resolutions.
Again, on a frame-by-frame basis, there is no averaging approach for obtaining the spectral density. But from all the above, we learned that the gradients in the frequency domain are extremely small, i.e. it is not neccessary at all to have a mask with edges. The mask is much more like a blurry cloud. And there is one more important fact: The very right of the frequency domain deals with hard edges of the image. It is not reasonable to correct too much there. But the more left hand areas of the frequency domain represent larger distances between pixels - this is where motion blur affects the image mostly.
In few words: it should be sufficient to blur the obtained frequency representation of a frame massively (and I really mean massively) and to restrict it to the left hand part in order to get the "equalizer mask". Again, the "equalizer mask" needed to be a point-by-point-divisor in the frequency domain.

Of course I'm also still eager to see the "globally subtract a little bit from the luma spectrum" feature in FFTQuiver() or DeFreq(). I experimented a bit with the script from the previous post and a chain of blur()'s instead of the motion blur, and it seems to me that it would be useful to have DeFreq() search for the point in frequency domain where there is the least spectral density instead of the most, and then subtract this amount from all frequencies before retransforming. It should help reduce the noise from those tiny digital camera sensors.

(to others who might come across this topic: don't forget to read the thread about the other capable FFT filter FFTQuiver from vcmohan: http://forum.doom9.org/showthread.php?t=107725

martin

[Mounir]

How does one use the show & info parameters?

I tried this for example Return DeFreq(fx=68,fy=-75,show=true,info=true) , but i get an error so i assume i'm doing something wrong.

[Fizick]

Mounir,
are your full script and "error" top secret ? ;)

[fbs]

I wish it wasn't dead.. because it somewhat works.. and it's the only filter I know which can do frequency removal..

ps.: I really didn't understand it's readme.. most stuff doesn't make sense

[tormento]

I someone has the source code, a modern implementation would be really useful, as I don't see any other frequency cleaner around.

If someone made up a more efficient plugin/script, please notice me.

[StainlessS]

I believe V.C.Mohan made a filter that is supposed to do similar.
His page on Wiki:- http://www.avisynth.nl/users/vcmohan/

a 5 second glance shows "vcm" plugin to be related.

This is a consolidation of vcmod, vcfreq, vcmove and addition of a few other utility functions. Some require libfftw3f-3.dll

EDIT:
FanFilter: Regular vertical frequency interference is filtered in spatial domain

FFTQuiver: 4 Functions for frequency domain customizable filtering in one or two dimensions. Homomorphic filtering can also be done. Requireslibfftw3f-3.dll or FFTw3.dll or fftw.dll to be in the path.

FQuiver: 2 Functions for frequency domain filtering in one or two dimensions, similar to FFTQuiver above. Does not require any other dll, but slower than FFTQuiver.

[tormento]

a 5 second glance shows "vcm" plugin to be related.
VCM is made for VS. AVS+ plugin is ManyPlus where the frequency manipulation has been split into multiple functions that makes me feel a bit dizzy.

[StainlessS]

Yes of course, sorry.