flawed megapixel experiment

acl wrote:
Ray Fischer wrote:

An example. Here's a ASCII representation of the light falling on
just one sensor in the array. I've divided the light up into 9 parts,
but the reality is there the number of discrete coloars is limited
only by the lens. The letters represent the obvious colors.

RRR
RGG
RRR

If the light-sensitive of the chip is just in the middle, then the
color it will detect is green. The other colors will be completely
invisible. Repeat that over many sensors and you get aliasing.


Wow! OK, look at this example
http://www.dpreview.com/learn/?/key=moire
The building has no red or blue (on the right photo), but the moire
pattern has these colours. So it can't be what you say, can it?

Sure can.

Well, unless you believe that "white is made up of all colours" means
that if I could sample a small enough patch of white light I'd get eg
red.

If you knew something about how the camera's sensor worked then the
explanation would be obvious. Apparently my simplified example has
only confused you.

Oh well.

--
Ray Fischer

Ray Fischer wrote:
acl wrote:
Ray Fischer wrote:

An example. Here's a ASCII representation of the light falling on
just one sensor in the array. I've divided the light up into 9 parts,
but the reality is there the number of discrete coloars is limited
only by the lens. The letters represent the obvious colors.

RRR
RGG
RRR

If the light-sensitive of the chip is just in the middle, then the
color it will detect is green. The other colors will be completely
invisible. Repeat that over many sensors and you get aliasing.


Wow! OK, look at this example
http://www.dpreview.com/learn/?/key=moire
The building has no red or blue (on the right photo), but the moire
pattern has these colours. So it can't be what you say, can it?

Sure can.

Well, unless you believe that "white is made up of all colours" means
that if I could sample a small enough patch of white light I'd get eg
red.

If you knew something about how the camera's sensor worked then the
explanation would be obvious. Apparently my simplified example has
only confused you.

Oh well.

OK then, please explain what these patches that you showed represent.
Also, explain why a regularly repeating pattern of black/white lines
can give rise to a) black and white moire b) colour effects (such as
shown in the example I linked to earlier). If your explanation involves
cutoffs in (spatial) frequency space, please explain the link to the
patches that you showed. If not, I think we have found the problem here.

Ilya Zakharevich writes:

P.S. Let me recall that the *crucial* difference between AA in audio
and photo is the "length" of the filter.

In audio, a (digital) 20-step AAF filter is close to a norm.
While in photo, the AAF in cameras is not even a 1-step filter!

It is not even low-pass filter, its MTF is |cos(f/f0)|. It just
happens to kill *some* frequencies. Since even such a filter is
not simple to construct, AND it has significant aberrations, it
is a very questionable tradeoff.

If the amount of offset between the two images that the AA filter
creates in the X direction (ignore Y for the moment) is exactly 1 pixel,
then the first zero in its response is at exactly 0.5 cycles/pixel. So
with such a filter you know that any luminance information at exactly
0.5 cycles/pixel horizontal frequency will simply turn into grey before
reaching the sensor.

Meanwhile, when coloured light passes through the Bayer filter, it
modulates the signal from that scanline at 0.5 cycles/pixel (because
the filter alternates R/G or G/B in one row or one column). So, if you
look at the signal from that row of the sensor, you can filter out a
narrow band of frequencies around 0.5 cycles/pixel and be pretty sure
that they are due to the colour of light falling on the Bayer filter,
not luminance detail at the same frequency (because that was already
suppressed).

So the AA filter has the important job of filtering out the range of
luminance frequencies that would be mistaken as colour information, and
displayed as colour moire, not just the frequencies that would alias to
produce wrong-frequency luminance information.

Colour moire is *really* annoying, so perhaps suppressing it is more
important than just general aliasing reduction. Maybe that's the main
thing the AA filter does, in which case having a very deep notch at 0.5
cycles/pixel is exactly what you want.

Dave

acl wrote:
Ray Fischer wrote:
acl wrote:
Ray Fischer wrote:

An example. Here's a ASCII representation of the light falling on
just one sensor in the array. I've divided the light up into 9 parts,
but the reality is there the number of discrete coloars is limited
only by the lens. The letters represent the obvious colors.

RRR
RGG
RRR

If the light-sensitive of the chip is just in the middle, then the
color it will detect is green. The other colors will be completely
invisible. Repeat that over many sensors and you get aliasing.


Wow! OK, look at this example
http://www.dpreview.com/learn/?/key=moire
The building has no red or blue (on the right photo), but the moire
pattern has these colours. So it can't be what you say, can it?

Sure can.

Well, unless you believe that "white is made up of all colours" means
that if I could sample a small enough patch of white light I'd get eg
red.

If you knew something about how the camera's sensor worked then the
explanation would be obvious. Apparently my simplified example has
only confused you.

Oh well.

OK then, please explain what these patches that you showed represent.

Already done.

Also, explain why a regularly repeating pattern of black/white lines
can give rise to a) black and white moire b) colour effects (such as

No.

--
Ray Fischer

[A complimentary Cc of this posting was sent to
Dave Martindale
], who wrote in article :
If the amount of offset between the two images that the AA filter
creates in the X direction (ignore Y for the moment) is exactly 1 pixel,
then the first zero in its response is at exactly 0.5 cycles/pixel. So
with such a filter you know that any luminance information at exactly
0.5 cycles/pixel horizontal frequency will simply turn into grey before
reaching the sensor.

Meanwhile, when coloured light passes through the Bayer filter, it
modulates the signal from that scanline at 0.5 cycles/pixel (because
the filter alternates R/G or G/B in one row or one column). So, if you
look at the signal from that row of the sensor, you can filter out a
narrow band of frequencies around 0.5 cycles/pixel and be pretty sure
that they are due to the colour of light falling on the Bayer filter,
not luminance detail at the same frequency (because that was already
suppressed).

So the AA filter has the important job of filtering out the range of
luminance frequencies that would be mistaken as colour information, and
displayed as colour moire, not just the frequencies that would alias to
produce wrong-frequency luminance information.

In theory, this would be fine.

In practice, the "zero"[*] of the MTF of AAF is not at 0.5
cycles/pixel, but slightly higher (about 0.65 cycles/px), as the
(calculated) MTF curves of the sensor+lens show. I do not know what
makes the camera manufactures choose this particular design; however,
this shows that *in their opinions* some other factors are at least as
important as one which you pointed out.
[*] since MTF=|OTF|, and OTF is complex, real-life MTF does not has
zeros, but just some small values where OTF becomes close to 0.

Thanks for a nice argument (I did not hear this put out so nicely before),
Ilya