Deconvolution of Sensor Anti-Alias filter?

Archived from groups: rec.photo.digital (More info?)

<winhag@yahoo.com> wrote in message
news:1126298522.839835.130970@g44g2000cwa.googlegroups.com...
> Folks,
>
> Another 'do they do that and if not, why?' question.
> Canon's sensors have an anti-alias filter for good reason.
> Most everyone uses some sort of Unsharp-Mask to 'restore'
> the 'sharpness' to the somewhat soft images.
> Since the anti-alias filter is known and well defined,
> wouldn't it make more sense to use a true 'reconstruction filter' (vs.
> USM)
> when processing RAW files as is typically done in sampled data systems?
> I would think this could avoid some of the USM artifacts.
> Does anyone know if this is done or is practical?
>
> W
>
No, the best place to put a filter is before the data gets sampled. Analog
filters also don't present much of a computational load on the camera.
Deconvolution is a very different process.
Jim

Archived from groups: rec.photo.digital (More info?)

winhag@yahoo.com wrote:

> I may not have been clear. The optical anti-alias filter is before the
> sensor. What I am suggesting is that a software reconstruction filter
> be used on the RAW data to undue the in-band suppression caused by the
> optical anti-alias filter.

Makes sense to me. Possibly USM is 'good enough' that there isn't enough
interest in doing what you suggest. Especially as it seems to be the
fashion that the 'appropriate' amount of sharpening is 'as much as
you can possibly get away with without introducing gross artefacts'.

- Len

Archived from groups: rec.photo.digital (More info?)

<winhag@yahoo.com> wrote in message
news:1126298522.839835.130970@g44g2000cwa.googlegroups.com...
> Folks,
>
> Another 'do they do that and if not, why?' question.
> Canon's sensors have an anti-alias filter for good reason.
> Most everyone uses some sort of Unsharp-Mask to
> 'restore' the 'sharpness' to the somewhat soft images.
> Since the anti-alias filter is known and well defined,
> wouldn't it make more sense to use a true 'reconstruction
> filter' (vs. USM) when processing RAW files as is typically
> done in sampled data systems?

Yes, but you should ideally also include the interaction of the lens
with the AA-filter in the same equation.

> I would think this could avoid some of the USM artifacts.
> Does anyone know if this is done or is practical?

I do it,
<http://www.xs4all.nl/~bvdwolf/main/downloads/Batavia_Crop.jpg>
demonstrates capture sharpening only, and it's very practical for
special (low volume) cases but also computationally expensive (very
slow). It is however not a widespread practice outside
astronomical/scientific photography.

One could, as an alternative, create a high-pass 'capture-restoration'
filter based on the Point-Spread-Function (PSF) of the imaging chain,
and it will largely compensate for capture losses without introducing
halo, but it will enhance noise if used without an edge mask.

There are several programs that allow deconvolution sharpening,
including Photoshop CS2 (Smart Sharpen filter), and DxO, and a few
others as well. The success usually depends on a good description of
the system blur (PSF), which can be obtained with e.g. Imatest
(www.imatest.com) software.

Bart

Archived from groups: rec.photo.digital (More info?)

Looks interesting. I understand your point about including the whole
chain, but I would be happy to
'attack' the anti-alias filter only. Since this is a known quantity (at
least by Canon) it would be nice
if they supplied software to compensate for it. I will look into the
options you mentioned.

Thanks

Archived from groups: rec.photo.digital (More info?)

<winhag@yahoo.com> wrote in message
news:1126320696.345901.76950@o13g2000cwo.googlegroups.com...
SNIP
> I will look into the options you mentioned.

You may also want to check out Image Analyzer
<http://meesoft.logicnet.dk/>, a free image analyzer/editor
(unfortunately only 8-bit/channel) written to assist with several
scientific imaging tasks.

Look at the "Filters|Restoration by deconvolution ..." option.
Choosing a Gaussian blur model with 0.70 to 0.95 radius, will add
sharpness with each iteration (often maxing out at 6 or 7) while
mostly avoiding typical USM artifacts.

Bart

Archived from groups: rec.photo.digital (More info?)

Thanks. Do you have any reference for the AA filter characteristics?

Archived from groups: rec.photo.digital (More info?)

"Bart van der Wolf" <bvdwolf@no.spam> writes:

>The closest one may come to it is what Canon discloses:
>http://www.canon.com/technology/d35mm/01.html

>The "point blur' introduced by such a filter will usually spread the
>signal pixel signal over 2 pixels,

4 pixels, because there are layers that provide horizontal and vertical
separation. The illustration shows how a high-frequency dot image gets
spread into 4 dots.

>Given a choice
>of material (Lithium Niobate) "thickness", and layers' orientation,
>distance will determine the amount of actual spread.

I've been told that manufacturers treat this is a tunable parameter. If
the spread is exactly one pixel pitch, you get a null in the response of
the system right at the Nyquist frequency, which is theoretically good.
But if you reduce the spread to perhaps 0.8 pixels, you get slightly
sharper-looking images at the risk of having more aliasing.

Dave

Archived from groups: rec.photo.digital (More info?)

[A complimentary Cc of this posting was sent to
Dave Martindale
<davem@cs.ubc.ca>], who wrote in article <dh43ej$6ce$1@mughi.cs.ubc.ca>:
> >Given a choice
> >of material (Lithium Niobate) "thickness", and layers' orientation,
> >distance will determine the amount of actual spread.
>
> I've been told that manufacturers treat this is a tunable parameter. If
> the spread is exactly one pixel pitch, you get a null in the response of
> the system right at the Nyquist frequency, which is theoretically good.

I do not see how this may be "theoretically good". There is no
aliasing at the Nyquist frequency, it appears *above* Nyquist. So
having a zero at Nyquist has absolutely no point.

Extending this by continuity: if you have an image with visible
frequency at 1.1 Nyquist, you get it aliased to 0.9 Nyquist. My
impression is that such a small difference between "actual" and
"recorded" frequency will not create any *visible* moire.

The "annoying" Moire is when some frequency in the image is aliased
into a *LOW* frequency in the recorded data. E.g., if some frequency
is aliased into one 1.5x smaller (I think this is the practical limit
of visible moire). This give 1.2 Nyquist aliased into 0.8 Nyquist.

So one should not care *much* about efficiency of anti-aliasing filter
below the frequency of 1.2 Nyquist. So it looks like optimal position
for a zero of MTF is about 1.4 Nyquist (so the critical range 1.2--1.6
Nyquist is "well-covered" ).

Of course, the actual numbers I used are hunches only; it is easy to
simulate the process on computer and compare results *visually* to
obtain the "best" choices.

Hope this helps,
Ilya

Archived from groups: rec.photo.digital (More info?)

"Ilya Zakharevich" <nospam-abuse@ilyaz.org> wrote in message
newsh87rb$2t58$1@agate.berkeley.edu...
SNIP
> Extending this by continuity: if you have an image with visible
> frequency at 1.1 Nyquist, you get it aliased to 0.9 Nyquist. My
> impression is that such a small difference between "actual" and
> "recorded" frequency will not create any *visible* moire.

Also note that the Green sampling density is different from Red and
Blue (in an GRGB mosaic). This can lead to false color-moire unless
properly dealt with in the Raw conversion.

Bart

Archived from groups: rec.photo.digital (More info?)

<winhag@yahoo.com> wrote:
> Anyone venture a guess on why (I believe) Kodak does not put AA filters
> on their sensors a la the Hasselblad H1/H2 and
> (discontinued) Kodak SLR's?

A cynical marketing sleaze to fool mathematically naive photographers into
thinking their products are sharper than they really are.

David J. Littleboy
Tokyo, Japan

Archived from groups: rec.photo.digital (More info?)

Ilya Zakharevich <nospam-abuse@ilyaz.org> writes:

>> I've been told that manufacturers treat this is a tunable parameter. If
>> the spread is exactly one pixel pitch, you get a null in the response of
>> the system right at the Nyquist frequency, which is theoretically good.

>I do not see how this may be "theoretically good". There is no
>aliasing at the Nyquist frequency, it appears *above* Nyquist. So
>having a zero at Nyquist has absolutely no point.

More precisely (though you really ought to know this): the sampling
theorem says that the input frequencies must be strictly *less than* the
Nyquist limit of 0.5 cycles/pixel. Although input at exactly the
Nyquist frequency does not alias, it cannot be sampled reliably either.
The measured amplitude can be double the true amplitude (if the sample
points are all at peaks) or zero (if the sample points are all at zero
crossings) or anywhere in between, depending on the relative phase of
the signal and the sampling clock. So input at the Nyquist frequency is
supposed to be filtered before sampling.

>Extending this by continuity: if you have an image with visible
>frequency at 1.1 Nyquist, you get it aliased to 0.9 Nyquist. My
>impression is that such a small difference between "actual" and
>"recorded" frequency will not create any *visible* moire.

It may not be Moire, but it's still wrong. This is visible in
resolution tests of the Sigma SD-9, where the 9-line resolution target
appears to be "resolved" at 2000 lines per picture height (1000 lp/ph),
but there are only 5 bars. It's a lovely example of aliasing. I don't
want *my* camera doing this.

>The "annoying" Moire is when some frequency in the image is aliased
>into a *LOW* frequency in the recorded data. E.g., if some frequency
>is aliased into one 1.5x smaller (I think this is the practical limit
>of visible moire). This give 1.2 Nyquist aliased into 0.8 Nyquist.

Again, Moire isn't the only problem introduced by aliasing.

>Of course, the actual numbers I used are hunches only; it is easy to
>simulate the process on computer and compare results *visually* to
>obtain the "best" choices.

Though note that the "best" choice obtained this way depends on both the
subject material and the viewer. The advantage of the conservative
approach of putting the filter's null right at the Nyquist frequency is
that the image content is more often correct, even if it appears a bit
less sharp than when a higher-cutoff AA filter was used.

Dave

Archived from groups: rec.photo.digital (More info?)

winhag@yahoo.com wrote:
> Anyone venture a guess on why (I believe) Kodak does not put AA
> filters on their sensors a la the Hasselblad H1/H2 and
> (discontinued) Kodak SLR's?

- cost - leaving out the AA filter lowers cost

- they didn't think that the lenses used would be good enough to produce
significant information abouve Nyquist?

- "the pictures look better without" ?

- "the resolution is already low enough - don't reduce it further" ?

- they didn't realise one would be required? Nah!

David

Archived from groups: rec.photo.digital (More info?)

[A complimentary Cc of this posting was sent to
David J Taylor
<david-taylor@blueyonder.co.not-this-bit.nor-this-part.uk.invalid>], who wrote in article <Rm7_e.116477$G8.86155@text.news.blueyonder.co.uk>:
> >> It would be interesting to see the results of such tests. My hunch
> >> is that images from cameras with AA filters having a zero MTF just
> >> /before/ Nyquist would look subtly better than those where above
> >> Nyquist (e.g. 1.2) is allowed.

> > Why do you think so? E.g., consider 0.9 vs 1.2; consider a pattern
> > with frequency 1.5 Nyquist. Your choice will pass through 87% of the
> > aliased pattern; mine only 38%.

• Note that the aliased pattern

> > has frequency of 0.5 Nyquist, so should be very noticable.

> That's not what I meant - I meant a filter which had a cut-off of 0.9 of
> the sampling frequency. You were suggesting a cut-off just above, I'm
> suggesting keep it below as theory requires.

As I show above, the theory requires exactly the opposite. Or did I
misunderstand you? What is a "cut-off"? AAFs have no cut-off, their
MTF is given by cosine law. What I was discussing was the position of
the first zero.

Hope this helps,
Ilya

Archived from groups: rec.photo.digital (More info?)

Ilya Zakharevich <nospam-abuse@ilyaz.org> writes:

>> More precisely (though you really ought to know this): the sampling
>> theorem says that the input frequencies must be strictly *less than* the
>> Nyquist limit of 0.5 cycles/pixel. Although input at exactly the
>> Nyquist frequency does not alias, it cannot be sampled reliably either.

>This is irrelevant, since the image on the focal plane has a
>continuous Fourier spectrum, so each *individual* frequency comes with
>0 amplitude. [You can calculate the "signal power" inside any
>*region* of frequencies, but the power goes down when the region
>narrows.]

A regular pattern in object space produces some finite amplitude of a
particular set of frequencies in image space. The power is there no
matter how narrow you make the region of frequencies. So power does
*not* go to zero as you narrow the region, for a single-frequency
signal. It doesn't matter that the focal-plane image is continuous, or
that its Fourier transform is continuous.

But, in practical photography, any given frequency is either above or
below Nyquist - you can never hit exactly Nyquist. So we should not use
it in examples (like you did).

But if you want to be precise, if you managed to have image content at
exactly the Nyquist frequency, it would not be reliably sampled. That's
why the sampling theorem says "less than", not "less than or equal".

>> It may not be Moire, but it's still wrong. This is visible in
>> resolution tests of the Sigma SD-9, where the 9-line resolution target
>> appears to be "resolved" at 2000 lines per picture height (1000 lp/ph),
>> but there are only 5 bars. It's a lovely example of aliasing. I don't
>> want *my* camera doing this.

>Interesting; do you have an URL? Although my immediate suspicion
>would be the demosaicing software, not AAF.

Look at the dpreview review of the SD9. Find the resolution test chart,
and download the full-size image. Take a look at the resolution wedge
at its narrowest point.

Dave

Archived from groups: rec.photo.digital (More info?)

In article <dh87rb$2t58$1@agate.berkeley.edu>, Ilya Zakharevich
<nospam-abuse@ilyaz.org> writes
>[A complimentary Cc of this posting was sent to
>Dave Martindale
><davem@cs.ubc.ca>], who wrote in article <dh43ej$6ce$1@mughi.cs.ubc.ca>:
>> >Given a choice
>> >of material (Lithium Niobate) "thickness", and layers' orientation,
>> >distance will determine the amount of actual spread.
>>
>> I've been told that manufacturers treat this is a tunable parameter. If
>> the spread is exactly one pixel pitch, you get a null in the response of
>> the system right at the Nyquist frequency, which is theoretically good.
>
>I do not see how this may be "theoretically good". There is no
>aliasing at the Nyquist frequency, it appears *above* Nyquist. So
>having a zero at Nyquist has absolutely no point.
>
Its onset is infinitesimally above Nyquist, hence Nyquist is the
effective limit and is why the null should be there or below.

>Extending this by continuity: if you have an image with visible
>frequency at 1.1 Nyquist, you get it aliased to 0.9 Nyquist. My
>impression is that such a small difference between "actual" and
>"recorded" frequency will not create any *visible* moire.

Perhaps not moire, since that requires an extended regular spatial
frequency source, but it will have deleterious effects, such as making
specular reflections larger than they should be. In terms of direct
moire, misrepresenting 1.1x Nyquist as 0.9x Nyquist is an exaggeration
of spatial frequencies by 20%, which certainly is noticeable when a
dominant image component of that rate occurs in the image.

In my particular field, I have seen images of tank with 8 regularly
spaced wheels on their tracks appearing to have only 6 wheels and
therefore being identified as a different vehicle completely as a
consequence of less aliasing than the 20% example that you reference. If
you don't think that is significant then I guess you don't think "blue
on blue" attacks are significant either!
>
>The "annoying" Moire is when some frequency in the image is aliased
>into a *LOW* frequency in the recorded data.

Not at all - when a high spatial frequency is aliased to a low one the
image distortion is so obvious as not to be objectionable in many cases.
The really objectionable distortion is the one that is not immediately
obvious, but unintentionally misrepresents the image.

--
Kennedy
Yes, Socrates himself is particularly missed;
A lovely little thinker, but a bugger when he's pissed.
Python Philosophers (replace 'nospam' with 'kennedym' when replying)

Archived from groups: rec.photo.digital (More info?)

[A complimentary Cc of this posting was sent to
Kennedy McEwen
<rkm@kennedym.demon.co.uk>], who wrote in article <cSbpNzDIleODFwEO@kennedym.demon.co.uk>:

> >This is irrelevant, since the image on the focal plane has a
> >continuous Fourier spectrum, so each *individual* frequency comes with
> >0 amplitude. [You can calculate the "signal power" inside any
> >*region* of frequencies, but the power goes down when the region
> >narrows.]

> Sorry Ilya, but that is just BS.

I'm forced to return the compliment...

> It doesn't matter that the power in an infinitesimal spatial
> bandwidth tends to zero, the important parameter is the power
> density.

With finite power density, what happens at *one particular frequency*
does not matter. So a sampling theorem holding for "less than
Nyquist" implies that it also holds for "less or equal to Nyquist";
actually, for continuous spectrum there is no difference between these
two formulations.

Hope this helps,
Ilya

Archived from groups: rec.photo.digital (More info?)

[A complimentary Cc of this posting was sent to
Kennedy McEwen
<rkm@kennedym.demon.co.uk>], who wrote in article <jTt2tOEQleODFwCc@kennedym.demon.co.uk>:

[Some more thoughts on the same subject]

> >Extending this by continuity: if you have an image with visible
> >frequency at 1.1 Nyquist, you get it aliased to 0.9 Nyquist. My
> >impression is that such a small difference between "actual" and
> >"recorded" frequency will not create any *visible* moire.
>
> Perhaps not moire, since that requires an extended regular spatial
> frequency source, but it will have deleterious effects, such as making
> specular reflections larger than they should be. In terms of direct
> moire, misrepresenting 1.1x Nyquist as 0.9x Nyquist is an exaggeration
> of spatial frequencies by 20%, which certainly is noticeable when a
> dominant image component of that rate occurs in the image.

Even in applications where this is noticable (very few), note that the
AAF with zero at 1.2 Nyquist will perform 2.6x times better to
eliminate this particular example of aliasing.

> In my particular field, I have seen images of tank with 8 regularly
> spaced wheels on their tracks appearing to have only 6 wheels and
> therefore being identified as a different vehicle completely as a
> consequence of less aliasing than the 20% example that you
> reference.

Actually, this is yet more water on my wheel. ;-) Your example is an
example of aliasing 1.14 to 0.86. For this case, AAF with zero at 1.2
will perform 5.2x better than one with zero at 0.9 of Nyquist.

Even you compare an AAF with zero at 1.2 of Nyquist with one at
Nyquist, the former will perform 2.8x better.

Hope this helps,
Ilya

Archived from groups: rec.photo.digital (More info?)

Ilya Zakharevich wrote:
> [A complimentary Cc of this posting was sent to
> David J Taylor
> <david-taylor@blueyonder.co.not-this-bit.nor-this-part.uk.invalid>],
> who wrote in article
> <Rm7_e.116477$G8.86155@text.news.blueyonder.co.uk>:
>>>> It would be interesting to see the results of such tests. My hunch
>>>> is that images from cameras with AA filters having a zero MTF just
>>>> /before/ Nyquist would look subtly better than those where above
>>>> Nyquist (e.g. 1.2) is allowed.
>
>>> Why do you think so? E.g., consider 0.9 vs 1.2; consider a pattern
>>> with frequency 1.5 Nyquist. Your choice will pass through 87% of
>>> the aliased pattern; mine only 38%.

• Note that the aliased

>>> pattern has frequency of 0.5 Nyquist, so should be very noticable.
>
>> That's not what I meant - I meant a filter which had a cut-off of
>> 0.9 of the sampling frequency. You were suggesting a cut-off just
>> above, I'm suggesting keep it below as theory requires.
>
> As I show above, the theory requires exactly the opposite. Or did I
> misunderstand you? What is a "cut-off"? AAFs have no cut-off, their
> MTF is given by cosine law. What I was discussing was the position of
> the first zero.
>
> Hope this helps,
> Ilya

There's a factor of two missing here! My last message was confused, I
meant 0.9 of the Nyquist frequency - half the sampling frequency. Sorry!
I'm more used to electrical AA filters, where you aim to get a brickwall
response (whilst not compromising phase linearity too much).

I am not familiar with the expected response of optical AA filters. Can
you point me to a plot of what you mean by cosine response - I hope you
don't mean that if the zero is at half the sampling frequency, the
response at the sampling frequency is -1.

David

Archived from groups: rec.photo.digital (More info?)

[A complimentary Cc of this posting was sent to
David J Taylor
<david-taylor@blueyonder.co.not-this-bit.nor-this-part.uk.invalid>], who wrote in article <0dt_e.117155$G8.29444@text.news.blueyonder.co.uk>:

> I am not familiar with the expected response of optical AA filters. Can
> you point me to a plot of what you mean by cosine response - I hope you
> don't mean that if the zero is at half the sampling frequency, the
> response at the sampling frequency is -1.

Yes it is (but this is for the splitter which separates the (two)
images by 2 pixels). A splitter which separates the images by 1 pixel
has 0 response at Nyquist frequency, and response -1 at twice the
Nyquist.

All AAF does it breaks image into two, and moves the parts aside a
little bit (well, the actual AAF does it twice: horizontally and
vertically). If it moves it the distance A, then from signal 2f(X) you
get f(X + A/2) + f(X - A/2).

Hope this helps,
Ilya

• Again, this assumes completely in-coherent light. For

completely-coherent light you get the square of this. This is why my
question about length of coherence is relevant.

Archived from groups: rec.photo.digital (More info?)

[A complimentary Cc of this posting was sent to
Kennedy McEwen
<rkm@kennedym.demon.co.uk>], who wrote in article <FpJfjPCAjlODFw0p@kennedym.demon.co.uk>:
> >Exactly the opposite. As you know, stronger AAF blur stronger.

> However that blur of the AAF is simply the rejection of higher spatial
> frequencies. Aliasing is the misrepresentation of high spatial
> frequencies with lower ones. The AAF does not represent fine detail
> such as specular reflections with coarse detail - aliasing does.

To see the error in your arguments, put an AAF designed for sensor
with step 8microns on a sensor with step 2.2microns; make a shot of
the same subject with the same lens. Aliasing disappears. Blur
remains.

> >This may sound impressive, but we are talking about difference *much*
> >smaller than the pixel size.

> If you want an estimate of how bad that actually is, talk to any of the
> film scanner people about the effect known as grain aliasing - exactly

All discussion I saw about the so called "grain aliasing" is pure
uneducated guesses. They see some effect, introduce a fancy name for it,
and try to discuss as if they know what they are talking about. (Maybe
I saw wrong sites; if you know something better, I would be glad to
revisit the question.)

> the same effect, misrepresenting fine undersampled grain as coarse grain
> in the final image. The same effect occurs here - fine textural
> information being misrepresented as coarse texture with the same
> contrast.

Again, you stray from the topic: if fine structure is aliased into
coarse, we are not talking about aliasing *near Nyquist limit*.

> >> In my particular field, I have seen images of tank with 8 regularly
> >> spaced wheels on their tracks appearing to have only 6 wheels and
> >> therefore being identified as a different vehicle completely as a
> >> consequence of less aliasing than the 20% example that you reference.

> >Well, if you do not know your tools, errors are always possible.

> Which is why correct filtering of the sensor is important - the end user
> shouldn't need to understand the detailed operation of the system or be
> required to compensate for the tools errors.

This depends on the application area of a tool. Most tools are not
subject to this requirement.

=======================================================

Moreover, your arithmetic is way wrong: in your example (8 aliased to
6) is NOT "less aliasing than the 20% example that you reference". To
nitpick, my example is actually 22% aliasing. Yours is 33% aliasing
(using the same metric).

> >> If you don't think that is significant then I guess you don't
> >> think "blue on blue" attacks are significant either!

> >Since I do not know what you mean, I suppose that "I do not think
so" indeed.

> Misinterpretation of targets results in errors which can, and in many
> cases have, resulted in fratricide or engagement of innocent bystanders.
> Seeing the wrong number of wheels on a vehicle can be enough to cause it
> to be identified as foe instead of friend.

Some people are ready to solve problems of "social" nature by
"technical" means (like: just introduce cleverly designed taxes, and
people will become nicer to each other). Your example looks of
similar flavor. It is people who fire weapons, not cameras. By
changing a design of AAF you won't decrease fratricide (even if all of
it is caused by results of misreading an image); most you can do is
decrease one kind while increasing some other kind.

> >This depends much on what is the purpose of your image. If you want
> >to distinguish 152mm howitzer (sp?) from 156mm one, then your choice
> >of tools may be very different from that of a bird photographer.

> So it would be unimportant to a bird photographer if a seagull was
> misinterpreted to be an albatross? I don't think so!

Let us get more details: you mean 2-pixel wide image of a seagull, or what?

Hope this helps,
Ilya

Archived from groups: rec.photo.digital (More info?)

Could the deconvolution of this be implemented in some sort of
Photoshop 'Custom filter'?

Ilya Zakharevich wrote:
> [A complimentary Cc of this posting was sent to
> David J Taylor
> <david-taylor@blueyonder.co.not-this-bit.nor-this-part.uk.invalid>], who wrote in article <0dt_e.117155$G8.29444@text.news.blueyonder.co.uk>:
>
> > I am not familiar with the expected response of optical AA filters. Can
> > you point me to a plot of what you mean by cosine response - I hope you
> > don't mean that if the zero is at half the sampling frequency, the
> > response at the sampling frequency is -1.
>
> Yes it is (but this is for the splitter which separates the (two)
> images by 2 pixels). A splitter which separates the images by 1 pixel
> has 0 response at Nyquist frequency, and response -1 at twice the
> Nyquist.

>
> All AAF does it breaks image into two, and moves the parts aside a
> little bit (well, the actual AAF does it twice: horizontally and
> vertically). If it moves it the distance A, then from signal 2f(X) you
> get f(X + A/2) + f(X - A/2).
>
> Hope this helps,
> Ilya
>
>

• Again, this assumes completely in-coherent light. For

> completely-coherent light you get the square of this. This is why my
> question about length of coherence is relevant.

Archived from groups: rec.photo.digital (More info?)

In article <dhf996$1vhh$1@agate.berkeley.edu>, Ilya Zakharevich
<nospam-abuse@ilyaz.org> writes
>[A complimentary Cc of this posting was sent to
>Kennedy McEwen
><rkm@kennedym.demon.co.uk>], who wrote in article
><FpJfjPCAjlODFw0p@kennedym.demon.co.uk>:
>> >Exactly the opposite. As you know, stronger AAF blur stronger.
>
>> However that blur of the AAF is simply the rejection of higher spatial
>> frequencies. Aliasing is the misrepresentation of high spatial
>> frequencies with lower ones. The AAF does not represent fine detail
>> such as specular reflections with coarse detail - aliasing does.
>
>To see the error in your arguments, put an AAF designed for sensor
>with step 8microns on a sensor with step 2.2microns; make a shot of
>the same subject with the same lens. Aliasing disappears. Blur
>remains.
>
And this demonstrates what precisely? Of course blur remains because,
as I stated previously, the blur is caused by the attenuation of the
higher spatial frequencies in the image, NOT their misrepresentation at
lower spatial frequencies as occurs in aliasing.

>> If you want an estimate of how bad that actually is, talk to any of the
>> film scanner people about the effect known as grain aliasing - exactly
>
>All discussion I saw about the so called "grain aliasing" is pure
>uneducated guesses. They see some effect, introduce a fancy name for it,
>and try to discuss as if they know what they are talking about. (Maybe
>I saw wrong sites; if you know something better, I would be glad to
>revisit the question.)
>
Try sampling white noise - what happens? HF noise is aliased to low
frequencies. It occurs in every sampling system that is inadequately
filtered and film grain is a high frequency noise. Perhaps you did
visit the wrong sites, but it would appear that you are the person who
doesn't understand this.


>> the same effect, misrepresenting fine undersampled grain as coarse grain
>> in the final image. The same effect occurs here - fine textural
>> information being misrepresented as coarse texture with the same
>> contrast.
>
>Again, you stray from the topic: if fine structure is aliased into
>coarse, we are not talking about aliasing *near Nyquist limit*.
>
As I pointed out in the example you provided, the misrepresentation is
quite significant and can readily cause mis-identification of objects,
even quite near the Nyquist limit.
>
>Moreover, your arithmetic is way wrong: in your example (8 aliased to
>6) is NOT "less aliasing than the 20% example that you reference". To
>nitpick, my example is actually 22% aliasing. Yours is 33% aliasing
>(using the same metric).
>
My example was one that I have actually presented to over 100 trained
observers and consistently results in target misidentification. It is
at a level where problems can be guaranteed to occur, well beyond their
onset. Whilst 22% aliasing may result in less problems, even in
something as well defined as a military vehicle, misidentification will
certainly not be zero and is likely to be higher in objects (such as
birds) where the difference between species or gender is much less
dramatic.
>
>> So it would be unimportant to a bird photographer if a seagull was
>> misinterpreted to be an albatross? I don't think so!
>
>Let us get more details: you mean 2-pixel wide image of a seagull, or what?
>
Minimum criteria for 50% probability of identification of a well defined
object on static images is at least 12 pixels linearly (more correctly,
6 cycles resolved across the minimum dimension). This is a well known
standard that has been in use since the work of Johnson etc. in the
1940s and 50s. With more closely related objects or where a higher
probability is necessary the requirement can be more than 4-5x this.
--
Kennedy
Yes, Socrates himself is particularly missed;
A lovely little thinker, but a bugger when he's pissed.
Python Philosophers (replace 'nospam' with 'kennedym' when replying)

Archived from groups: rec.photo.digital (More info?)

Kennedy McEwen wrote:
[]
> Minimum criteria for 50% probability of identification of a well
> defined object on static images is at least 12 pixels linearly (more
> correctly, 6 cycles resolved across the minimum dimension). This is
> a well known standard that has been in use since the work of Johnson
> etc. in the 1940s and 50s. With more closely related objects or
> where a higher probability is necessary the requirement can be more
> than 4-5x this.

Remember as well that a well-trained observer (i.e. expert) can manager
with a picture that looks like a blur to us. Example: doctors examining
X-ray images. Also motion can play an important part - a human does not
move in the same way as a vehicle, so you don't need to see the limbs to
identify one versus the other.

David

Archived from groups: rec.photo.digital (More info?)

Ilya Zakharevich <nospam-abuse@ilyaz.org> writes:

>> I am not familiar with the expected response of optical AA filters. Can
>> you point me to a plot of what you mean by cosine response - I hope you
>> don't mean that if the zero is at half the sampling frequency, the
>> response at the sampling frequency is -1.

>Yes it is (but this is for the splitter which separates the (two)
>images by 2 pixels). A splitter which separates the images by 1 pixel
>has 0 response at Nyquist frequency, and response -1 at twice the
>Nyquist.

No, David was right - it's the response of a filter with one pixel
separation. That has a zero at the Nyquist frequency, which is half the
sampling frequency, and a -1 response at the sampling frequency which is
twice Nyquist.

Here, the -1 means that a signal at the sampling frequency is not
attenuated at all, but it is shifted half a pixel which is a 180 degree
phase shift which is equivalent to multiplying the signal by -1.

Dave

Archived from groups: rec.photo.digital (More info?)

In article <boO_e.117692$G8.15563@text.news.blueyonder.co.uk>, David J
Taylor
<david-taylor@blueyonder.co.not-this-bit.nor-this-part.uk.invalid>
writes
>Kennedy McEwen wrote:
>[]
>> Minimum criteria for 50% probability of identification of a well
>> defined object on static images is at least 12 pixels linearly (more
>> correctly, 6 cycles resolved across the minimum dimension). This is
>> a well known standard that has been in use since the work of Johnson
>> etc. in the 1940s and 50s. With more closely related objects or
>> where a higher probability is necessary the requirement can be more
>> than 4-5x this.
>
>Remember as well that a well-trained observer (i.e. expert) can manager
>with a picture that looks like a blur to us. Example: doctors examining
>X-ray images.

Yes, and these figure assume a trained expert - even more resolution is
required to enable a novice to achieve the same probability of
identification.

> Also motion can play an important part - a human does not
>move in the same way as a vehicle, so you don't need to see the limbs to
>identify one versus the other.

Yes, motion does play a part and the resolution requirements for moving
images are substantially lower than for still frames. However, we are
talking about DSLRs here, producing still frames. It doesn't actually
matter that you can identify an object in the viewfinder if you can't do
so in the resulting photograph.

"Honest Guvnor, that fuzzy blob is a Yeti, you could tell by the way it
moved" doesn't really cut much ice. ;-)
--
Kennedy
Yes, Socrates himself is particularly missed;
A lovely little thinker, but a bugger when he's pissed.
Python Philosophers (replace 'nospam' with 'kennedym' when replying)