Is this a famous filter, and how to compute the orthogonal filter?

Hi all:

A paper gives a filter:
  [ -2     3    -6     3    -2
     3     4     2     4     3
    -6     2    48     2    -6
     3     4     2     4     3
    -2     3    -6     3    -2 ]/64
and uses its orthogonal filter to solve a problem, but the author did
not give out what is its orthogonal filter exactly. I wonder is this
filter a famous one and the orthogonal filter is notorious?

I have not learned much about the filter. Any help will be
appreciated.

Zhang Zhiqiang

Zhang Zhiqiang wrote:
> Hi all:
>
> A paper gives a filter:
>   [ -2     3    -6     3    -2
>      3     4     2     4     3
>     -6     2    48     2    -6
>      3     4     2     4     3
>     -2     3    -6     3    -2 ]/64
> and uses its orthogonal filter to solve a problem, but the author did
> not give out what is its orthogonal filter exactly. I wonder is this
> filter a famous one and the orthogonal filter is notorious?
>
> I have not learned much about the filter. Any help will be
> appreciated.
>
> Zhang Zhiqiang

Hello Zhang,

I don't recognize this as a famous filter - I'm not sure if there are
really any famous FIR filters. Except for some windows.

However, just looking at the coefs, I see several things. First is the
even symmetry about the central value of 48, so this is a linear phase
filter. The sum of the coefs is 1 (after /64), so the DC response is
1.0. You have 25 coefs, therefore the filter's delay is 13 smaples. I
don't have access to a math package right now, but just find the
fourier transform of the 25 values and see what the frequency response
is. You can then derive the orthogonal filter from the FT of the filter
easily.

Later,
Clay

Clay wrote:
> Zhang Zhiqiang wrote:
> 
>>Hi all:
>>
>>A paper gives a filter:
>>  [ -2     3    -6     3    -2
>>     3     4     2     4     3
>>    -6     2    48     2    -6
>>     3     4     2     4     3
>>    -2     3    -6     3    -2 ]/64
>>and uses its orthogonal filter to solve a problem, but the author did
>>not give out what is its orthogonal filter exactly. I wonder is this
>>filter a famous one and the orthogonal filter is notorious?
>>
>>I have not learned much about the filter. Any help will be
>>appreciated.
>>
>>Zhang Zhiqiang
> 
> 
> Hello Zhang,
> 
> I don't recognize this as a famous filter - I'm not sure if there are
> really any famous FIR filters. Except for some windows.
> 
> However, just looking at the coefs, I see several things. First is the
> even symmetry about the central value of 48, so this is a linear phase
> filter. The sum of the coefs is 1 (after /64), so the DC response is
> 1.0. You have 25 coefs, therefore the filter's delay is 13 smaples. I
> don't have access to a math package right now, but just find the
> fourier transform of the 25 values and see what the frequency response
> is. You can then derive the orthogonal filter from the FT of the filter
> easily.
> 
> Later,
> Clay

Clay,

It looks like a 2D filter kernel to me. Is that consistent with what you 
wrote above? Note that the rows and columns are individually symmetric, 
and that there is symmetry about the diagonals. I have no idea what an 
orthogonal spatial filter is or what it might do, but I guess that this 
is an image-processing question.

Jerry
-- 
Engineering is the art of making what you want from things you can get.
�����������������������������������������������������������������������

Jerry Avins wrote:
>>
> It looks like a 2D filter kernel to me. Is that consistent with what
you
> wrote above? Note that the rows and columns are individually
symmetric,
> and that there is symmetry about the diagonals. I have no idea what
an
> orthogonal spatial filter is or what it might do, but I guess that
this
> is an image-processing question.
>
> Jerry
> --
> Engineering is the art of making what you want from things you can
get.
>
=AF=AF=AF=AF=AF=AF=AF=AF=AF=AF=AF=AF=AF=AF=AF=AF=AF=AF=AF=AF=AF=AF=AF=AF=AF=
=AF=AF=AF=AF=AF=AF=AF=AF=AF=AF=AF=AF=AF=AF=AF=AF=AF=AF=AF=AF=AF=AF=AF=AF=AF=
=AF=AF=AF=AF=AF=AF=AF=AF=AF=AF=AF=AF=AF=AF=AF=AF=AF=AF=AF=AF=AF


Hello Jerry,

My original thoughts were for a 1-D filter, but the even symmetry -
linear phase and DC response statements still hold for a 2-D cased. The
delay is now thought of as an offset to the center of the convolver
block.

An orthogonal filter block can be useful if one needs to know the real
intensity at a pixel's location. Or if one needs to fill in an unknown
pixel.

He can do a 2-D fourier transform, zero out the negative frequencies
and Inverse FT the data, and finally take the imaginary part for his
orthogonal filter. I didn't really answer his question earlier on how
to find an orthogonal filter. When I get back to a computer with a math
package, I'll check out the frequency response. This may be some kind
of interpolator used for demosaicing (I know that's an ugly word, but
it is what is commonly used now by people in that arena.)

Clay

"Zhang Zhiqiang" <mathzqy@gmail.com> wrote in message 
news:7348e98b.0503140523.7f92dde0@posting.google.com...
> Hi all:
>
> A paper gives a filter:
>  [ -2     3    -6     3    -2
>     3     4     2     4     3
>    -6     2    48     2    -6
>     3     4     2     4     3
>    -2     3    -6     3    -2 ]/64
> and uses its orthogonal filter to solve a problem, but the author did
> not give out what is its orthogonal filter exactly. I wonder is this
> filter a famous one and the orthogonal filter is notorious?
>
> I have not learned much about the filter. Any help will be
> appreciated.
>
> Zhang Zhiqiang

Zhang,

I'd never heard of it.  A quick Google revealed:
http://www.cs.sunysb.edu/~mueller/teaching/cse616/wavelets.ppt#54
where it says the orthogonal filter [presumably to h(t)] is:

g(k)=[(-1)^k]*h(N-k)

Also,

http://www.site.uottawa.ca/~edubois/courses/ELG5378/Two_channel_FB.pdf

gives some insight.

Some nice pictures at:

http://www.toolsmiths.com/fwrorth.htm

and so on.....

Fred

Fred 

"Clay" <physics@bellsouth.net> wrote in message news:<1110816207.551392.145950@l41g2000cwc.googlegroups.com>...
> Hello Jerry,
> 
> My original thoughts were for a 1-D filter, but the even symmetry -
> linear phase and DC response statements still hold for a 2-D cased. The
> delay is now thought of as an offset to the center of the convolver
> block.
> 
> An orthogonal filter block can be useful if one needs to know the real
> intensity at a pixel's location. Or if one needs to fill in an unknown
> pixel.
> 
> He can do a 2-D fourier transform, zero out the negative frequencies
> and Inverse FT the data, and finally take the imaginary part for his
> orthogonal filter. I didn't really answer his question earlier on how
> to find an orthogonal filter. When I get back to a computer with a math
> package, I'll check out the frequency response. This may be some kind
> of interpolator used for demosaicing (I know that's an ugly word, but
> it is what is commonly used now by people in that arena.)
> 
> Clay


Hi, Clay:

Thanks for your help. You are genius and have guessed that the filter
was used for demosaicing. Yes, that filter is from the paper "Color
demosaicing by estimating luminance and opponent chromatic signals in
the Fourier domain", which we can find it at
http://ivrgwww.epfl.ch/publications/ash02.pdf
and it is a 2-D filter.

Following your advice, if matlab is at hand, I first get the FFT
transform by command "fft_H = fft2(H)", but what is the meaning "the
negative frequencies"? Now, fft_H is a 5 by 5 matrix. Maybe I should
use the command "fftshift(fft_H)" to shift zero-frequency component of
discrete Fourier transform to center of spectrum. Then the negative
frequencies are fft_H([1 2], [1 2]), like the zero positions at the
matrix given below?
0 0 1 1 1;
0 0 1 1 1;
1 1 1 1 1;
1 1 1 1 1;
1 1 1 1 1;

Zhang Zhiqiang