Entry level Nikon 24mp?!

In article , Eric Stevens
wrote:

Lets start again.

ok!

One pixel is a cell collecting light.

yes, but it's actually a sensel. the terms are often confused.

One Bayer cell consists of 4 pixels: Red, Blue and 2 x Green.

wrong.

one bayer pixel is a matrix of multiple sensels which is often a 3x3
block of a given sensel plus its 8 surrounding sensels. there can be
more sensels used in the demosaicing too.

it's a lot more complex than people think it is, which is why there's
so much confusion.

Some people call a Bayer Cell a pixel. Other people call the basic
light collecting elements 'pixels'.

they can call it whatever they want, but a 2x2 block is not how bayer
works.

That four to one ratio is what lies behind my statement to which you
object.

it's wrong.

In article , Eric Stevens
wrote:

Four pixels make a Bayer cell.

no they definitely do not.

2.7 million Bayer cells = 2.7 x 4 million pixels = 10.8Mp

It all depends on how you define cell and pixel.

no it doesn't.

Wrong. :-(

it's not wrong.

'tis.

no it's not wrong. you're confused.

In article , Eric Stevens
wrote:

I suspect from what you have said that every active light-sensitive
pixel was surrounded by a 'binned' inactive pixel. The result would be
an array of 4 x 4 pixels such as:

X R X G
X X X X
X G X B
X X X X

That would give 4 active pixels for every 16.

that's not how they did it.

all they did was take a sensor that had 10.8 megapixels and cover every
2x2 block of sensels with a red, green or blue filter plus a microlens,
rather than cover a single sensel with a filter/microlens, as in a more
typical implementation.

instead of:
r g r g
g b g b
r g r g

you had:
r r g g r r g g
r r g g r r g g
g g b b g g b b
g g b b g g b b
r r g g r r g g
r r g g r r g g

there used to be a closeup picture of the sensor where you could see
the individual sensels but it's gone.

Eric Stevens wrote:
On Sat, 14 Apr 2012 02:09:29 -0800, (Floyd L.
Davidson) wrote:

Eric Stevens wrote:
rOn Tue, 10 Apr 2012 17:33:35 +1000, "Trevor" wrote:
"Eric Stevens" wrote:
The Nikon D1 was 2.4 Mp.

actually it was 2.7 mp from a 10.8 mp sensor.

That is absolutely correct.

From http://imaging.nikon.com/history/scenes/12/

"I guess that it's now safe to reveal that the D1 image
sensor, with specifications noting a pixel count of
2.7-million pixels, actually had a pixel count of 10.8-million
pixels. The technical reason for an actual pixel count four
times greater than that indicated publicly lies in the need to
achieve high sensitivity and a good signal-to-noise ratio.
Unlike current cameras, for which final pixel counts account
for individual pixels, we had to include multiple pixels in
each pixel unit with the D1." -- Kiyoshige Shibazaki, Nikon


Neither of us are right according to
http://www.nikonusa.com/Nikon-Produc...Tabs-TechSpecs


Says 2.7 Mp total, 2.66 Mp effective. Where do you get a 10.8Mp sensor?


Four pixels make a Bayer cell.

Balderdash. The Bayer filter has nothing to do with it. The
RAW file output from the camera is Bayer encoded, and has
2,663,888 pixel data values (2012x1324).

The actual sensor has 10,655,552 sensor locations, and 4 each of
them are binned in order to produce a single value for the Bayer
Encoded RAW file. The binning is done in hardware.

2.7 million Bayer cells = 2.7 x 4 million pixels = 10.8Mp

It all depends on how you define cell and pixel.

No it doesn't. First, the idea that a "Bayer cell" is 4 sensor
locations is ignorant. The *minimum* number of sensor locations
used to generate an image pixel is 4 (a 2x2 matrix), but in fact
all but the most basic raw conversion program will use either a
3x3 matrix or a 5x5 matrix. Hence pixels are usually generated
from either 9 sensor locations or 25, not 4.

Regardless the RAW data from the Nikon D1 contained only 2.66
MP. And after the Bayer CFA is decoded Nikon generated a
2000x3012 pixel image (and some raw converters used the entire
2012x3024).

Incidentally, I have a very functional Nikon D1 sitting within
arms reach.

And also, note that the D1X generated pixels by binning two
side by side sensor sites per pixel, rather than 4.

Thanks for all that. It makes sense of what was rapidly growing into a
directionless argument.

I suspect from what you have said that every active light-sensitive
pixel was surrounded by a 'binned' inactive pixel. The result would be
an array of 4 x 4 pixels such as:

X R X G
X X X X
X G X B
X X X X

That would give 4 active pixels for every 16.

Do you understand what "binning" is????

These arrays would be stacked to give:

^
|
X R X G X R X G
X X X X X X X X
X G X B X G X B
X X X X X X X X ---
X R X G X R X G
X X X X X X X X
X G X B X G X B
X X X X X X X X

I don't know why they did that, although I can think of a number of
possible reasons.

Of course they *didn't* do that.

See the reply from nospam, he explained it exactly.

--
Floyd L. Davidson http://www.apaflo.com/
Ukpeagvik (Barrow, Alaska)

On Sun, 15 Apr 2012 23:47:11 -0400, nospam
wrote:

In article , Eric Stevens
wrote:

I suspect from what you have said that every active light-sensitive
pixel was surrounded by a 'binned' inactive pixel. The result would be
an array of 4 x 4 pixels such as:

X R X G
X X X X
X G X B
X X X X

That would give 4 active pixels for every 16.

that's not how they did it.

all they did was take a sensor that had 10.8 megapixels and cover every
2x2 block of sensels with a red, green or blue filter plus a microlens,
rather than cover a single sensel with a filter/microlens, as in a more
typical implementation.

instead of:
r g r g
g b g b
r g r g

you had:
r r g g r r g g
r r g g r r g g
g g b b g g b b
g g b b g g b b
r r g g r r g g
r r g g r r g g

there used to be a closeup picture of the sensor where you could see
the individual sensels but it's gone.

Right. I get it.

Can you now explain what Floyd meant by 'binned'?

Regards,

Eric Stevens

Eric Stevens wrote:

Can you now explain what Floyd meant by 'binned'?

http://en.wikipedia.org/wiki/Data_binning

It's badly explained, unfortunately.

Basically, what you do is trade off resolution for sensitivity
and less read noise.

Instead of passing every pixels' captured electrons to the A/D
converter singly, you combine the charge of the pixels you are
binning and send the combined charge to the A/D converter.

If you bin 2x2, you have about[1] 4 times the electrons to
convert to a DN (digital number), which means a lot when the A/D
converter needs relatively many electrons compared to what your
pixel captures. Additionally, if you add them up now (instead
of after the A/D converter, e.g. by downsampling), you get the
read noise only once, instead of 4 times, which means SQRT(4) =
2 times less read noise.

Classic bayer pattern sensors are very hard to bin, since you'd have
to bin non-adjacent cells (so just transferring the charge to
the next cell in a CCD doesn't work) and you get strange
artifacts because you combine a rather large area, so with
bayer pattern sensors you will usually just downscale.

(Note that a
r r g g
r r g g
g g b b
g g b b
pattern isn't a classic bayer pattern unless you do bin it
properly 2x2.)

Binning is therefore usually found in scientific sensors, which
are often monochromatic and get their colour information (if any)
often from a colour wheel --- which will often not just include
red, green and blue, but, depending on the task, also other
bandpass, narrow bandpass and IR or UV filters.

Binning also reduces the amount of data created, which, when it
has to be transmitted, can be a problem when you're a million
miles from Earth.

http://www.ccd.com/ccd103.html
http://www.andor.com/learning/digita...ras/?docid=320
http://www.starrywonders.com/binning.html
http://www.noao.edu/outreach/aop/glossary/binning.html
http://www.photometrics.com/resource...ne/binning.php
http://www.roperscientific.de/binning.html


-Wolfgang

[1] photon noise!

Floyd L. Davidson wrote:
Paul Furman wrote:
Floyd L. Davidson wrote:
Eric wrote:
rOn Tue, 10 Apr 2012 17:33:35 +1000, wrote:
"Eric wrote:
The Nikon D1 was 2.4 Mp.

The actual sensor has 10,655,552 sensor locations, and 4 each of
them are binned in order to produce a single value for the Bayer
Encoded RAW file. The binning is done in hardware.

They should do that with any new 24MP camera, as an option at least.

It's probably not possible implement as a switchable option in
hardware without introducing serious noise.

Noise is not the problem. Electronics, not being a CCD and
having to bin over large spaces (to bridge between 4 reds or
blues) is. Image quality is too, then.

Some Canons have a smaller optional raw format.

Canon's "smaller option raw format" is not a raw format.

It is a raw format, but not in the sense that you get the
direct output of each sensel. In practice those that use it,
the advantages they use RAW for, but without having to store
all the data they don't care for in first place.

What's not clear to me, is whether that makes a sharper low res image in
the same way reducing after raw conversion will do.

Resizing to a smaller resolution does not make an image
"sharper" as such.

It thins the borders between ajdacent areas, just as wavelet
sharpen does. So not only the acutancy can rise.

It removes high frequency detail,

which, depending on the image and the amount of downsizing,
isn't there in first place (excluding high frequency noise).
From bayer patterns being very good but not perfect in the
restoration of high frequencies ( SQRT(2) of the pixel to pixel
distance) over the roll-off of AA filters down to camera shake,
slight misfocus and subject movement ... there can be many reasons
why high frequency detail simply isn't there in first place.

and in
that sense reduces sharpness.

Which is irrelevant, unless you're actually seeing it at a
distance and resolution where such detail, if it was were,
would be visible. Most images just aren't viewed at 100%.

Most photos aren't
all that sharp at full resolution, if for no other reason than the
existence of an antialiasing filter.

That isn't really true. The AA filter, in a properly processed
image, has just about exactly the same amount of high frequency
detail as a similar camera without the AA filter, except that at
frequencies very close to the Nyquist limit the signal to noise
ratio will be slightly reduced on the camera with the AA filter
(and conversely on camera without the AA filter the SNR will be
reduced by aliasing distortion throughout the frequency
spectrum).

That assumes a hard cut-off of the AA filter. I understand
that that's pretty hard to do in the real world.

In any case the reason images are apparently not sharp when
viewed at 100% is because a Bayer Color Filter encoded camera
simply cannot produce a tone transition in less than some number
of pixels, and the larger the demosiacing matrix the higher the
minimum for transition, as well as the more accurate the colors.

Actually, you are wrong here. Intelligent demosaicing
detects borders and doesn't smear them, irrespective of their
size. It's not a simple averaging process.

Of course that isn't really "sharpness", but acutance, and again
it can be increased with either a high pass filter (Sharpen) or
application of Unsharp Mask.

Of course that's really sharpness, as such an averaging process
would reduce or destroy the visibility of e.g. thin parallel
lines, thus reducing resolution drastically.

-Wolfgang

Wolfgang Weisselberg wrote:
Floyd L. Davidson wrote:
Paul Furman wrote:
Floyd L. Davidson wrote:
Eric wrote:
rOn Tue, 10 Apr 2012 17:33:35 +1000, wrote:
"Eric wrote:
The Nikon D1 was 2.4 Mp.

The actual sensor has 10,655,552 sensor locations, and 4 each of
them are binned in order to produce a single value for the Bayer
Encoded RAW file. The binning is done in hardware.

They should do that with any new 24MP camera, as an option at least.

It's probably not possible implement as a switchable option in
hardware without introducing serious noise.

Noise is not the problem. Electronics, not being a CCD and
having to bin over large spaces (to bridge between 4 reds or
blues) is. Image quality is too, then.

Yes. That adds up to *noise* is the problem.

Some Canons have a smaller optional raw format.

Canon's "smaller option raw format" is not a raw format.

It is a raw format, but not in the sense that you get the
direct output of each sensel.

It simply is not a raw format. It is interpolated data, not
raw data that needs to be interpolated.

In practice those that use it,
the advantages they use RAW for, but without having to store
all the data they don't care for in first place.

It's just a TIFF RGB format. Whoopee. They could shoot
JPEG to get the same effect...

What's not clear to me, is whether that makes a sharper low res image in
the same way reducing after raw conversion will do.

Resizing to a smaller resolution does not make an image
"sharper" as such.

It thins the borders between ajdacent areas, just as wavelet
sharpen does. So not only the acutancy can rise.

It removes high frequency detail,

which, depending on the image and the amount of downsizing,
isn't there in first place (excluding high frequency noise).

Well, its true that if you shoot pictures of gray cards, there
isn't much high frequency detail. Otherwise there is.

From bayer patterns being very good but not perfect in the
restoration of high frequencies ( SQRT(2) of the pixel to pixel
distance) over the roll-off of AA filters down to camera shake,
slight misfocus and subject movement ... there can be many reasons
why high frequency detail simply isn't there in first place.

But for good photographers, who use good technique, there is
almost always enough high frequency detail to make a very
visible difference.

That is *precisely* why applying Sharpen almost always has a
significant effect.

and in
that sense reduces sharpness.

Which is irrelevant, unless you're actually seeing it at a
distance and resolution where such detail, if it was were,
would be visible. Most images just aren't viewed at 100%.

Prints commonly are.

Most photos aren't
all that sharp at full resolution, if for no other reason than the
existence of an antialiasing filter.

That isn't really true. The AA filter, in a properly processed
image, has just about exactly the same amount of high frequency
detail as a similar camera without the AA filter, except that at
frequencies very close to the Nyquist limit the signal to noise
ratio will be slightly reduced on the camera with the AA filter
(and conversely on camera without the AA filter the SNR will be
reduced by aliasing distortion throughout the frequency
spectrum).

That assumes a hard cut-off of the AA filter. I understand
that that's pretty hard to do in the real world.

It is virtually impossible to do, but that is not required
for what was described.

The high frequency components just below the Nyquist Limit are
reduced by the AA filter, but by no means eliminated because the
filter is not particularly sharp. Because of that the actual
frequency where maximum frequency distortion occurs will
intentionally be placed just above the Nyquist Limit. Most
designs will hit a minimum and at higher frequencies will have a
slope similar to the slope at lower frequencies. (On some that
slope may be very steep at frequencies well above the Nyquist
Limit, but hopefully that will also be at the upper limits for
any given lens that might be used, so between the filter and the
lens there is a very low response level at those frequencies
too.)

Whatever, the AA filter is a Low Pass filter, and a High Pass
filter can be applied in post processing with an almost opposite
frequency response to perfectly correct the rolloff provided
by the AA filter. At frequencies just below the Nyquist Limit (and
be aware that here are absolutely *no* frequencies above the Limit) the
correction from the Sharpen HP filter is at its highest. The filter
increases noise just as much as it increases desired signal, and
that is why the SNR is worse than it would without the AA filter but
the frequency response will be almost precisely the same.

Of course at lower and lower frequencies there is less and less
correction by the Sharpen filter, and therefore less and less
change to the SNR compared to an image taken without the AA filter.

In any case the reason images are apparently not sharp when
viewed at 100% is because a Bayer Color Filter encoded camera
simply cannot produce a tone transition in less than some number
of pixels, and the larger the demosiacing matrix the higher the
minimum for transition, as well as the more accurate the colors.

Actually, you are wrong here. Intelligent demosaicing
detects borders and doesn't smear them, irrespective of their
size. It's not a simple averaging process.

You simply cannot get a one pixel length tone transition,
without applying a Sharpen HP or USM filter. Yes different
demosaicing algorithms can produce sharper results, but absent
some form of sharpen, there are no truly "sharp" transitions.

Of course that isn't really "sharpness", but acutance, and again
it can be increased with either a high pass filter (Sharpen) or
application of Unsharp Mask.

Of course that's really sharpness, as such an averaging process
would reduce or destroy the visibility of e.g. thin parallel
lines, thus reducing resolution drastically.

There is no actual increase in resolution. All that happens
with either USM or an HP filter is that the difference between
existing tone transition edges in increased.

--
Floyd L. Davidson http://www.apaflo.com/
Ukpeagvik (Barrow, Alaska)

On Mon, 16 Apr 2012 13:01:21 +0200, Wolfgang Weisselberg
wrote:

Eric Stevens wrote:

Can you now explain what Floyd meant by 'binned'?

http://en.wikipedia.org/wiki/Data_binning

It's badly explained, unfortunately.

Basically, what you do is trade off resolution for sensitivity
and less read noise.

Instead of passing every pixels' captured electrons to the A/D
converter singly, you combine the charge of the pixels you are
binning and send the combined charge to the A/D converter.

If you bin 2x2, you have about[1] 4 times the electrons to
convert to a DN (digital number), which means a lot when the A/D
converter needs relatively many electrons compared to what your
pixel captures. Additionally, if you add them up now (instead
of after the A/D converter, e.g. by downsampling), you get the
read noise only once, instead of 4 times, which means SQRT(4) =
2 times less read noise.

Right. I get it. That's the statistical use of the term 'binned'. The
problem I had is that there are several different
meanings/applications for the term 'binned' and none of them seemed to
make sense in the particular context. The closest I could get was
'binned' as in 'dumped' i.e. the data was ignored.

Classic bayer pattern sensors are very hard to bin, since you'd have
to bin non-adjacent cells (so just transferring the charge to
the next cell in a CCD doesn't work) and you get strange
artifacts because you combine a rather large area, so with
bayer pattern sensors you will usually just downscale.

Understood.

(Note that a
r r g g
r r g g
g g b b
g g b b
pattern isn't a classic bayer pattern unless you do bin it
properly 2x2.)

Binning is therefore usually found in scientific sensors, which
are often monochromatic and get their colour information (if any)
often from a colour wheel --- which will often not just include
red, green and blue, but, depending on the task, also other
bandpass, narrow bandpass and IR or UV filters.

Binning also reduces the amount of data created, which, when it
has to be transmitted, can be a problem when you're a million
miles from Earth.

http://www.ccd.com/ccd103.html
http://www.andor.com/learning/digita...ras/?docid=320
http://www.starrywonders.com/binning.html
http://www.noao.edu/outreach/aop/glossary/binning.html
http://www.photometrics.com/resource...ne/binning.php
http://www.roperscientific.de/binning.html


-Wolfgang

[1] photon noise!

Many thanks.

Regards,

Eric Stevens

Floyd L. Davidson wrote:
Wolfgang Weisselberg wrote:
Floyd L. Davidson wrote:
Paul Furman wrote:
Floyd L. Davidson wrote:
Eric wrote:
rOn Tue, 10 Apr 2012 17:33:35 +1000, wrote:
"Eric wrote:
The Nikon D1 was 2.4 Mp.

The actual sensor has 10,655,552 sensor locations, and 4 each of
them are binned in order to produce a single value for the Bayer
Encoded RAW file. The binning is done in hardware.

They should do that with any new 24MP camera, as an option at least.

It's probably not possible implement as a switchable option in
hardware without introducing serious noise.

Noise is not the problem. Electronics, not being a CCD and
having to bin over large spaces (to bridge between 4 reds or
blues) is. Image quality is too, then.

Yes. That adds up to *noise* is the problem.

That's like reasoning that drugs are the problem
for accidents on icy roads.

At least *try* to understand that not all image problems are
noise and that not all technological problems result in noise.


Some Canons have a smaller optional raw format.

Canon's "smaller option raw format" is not a raw format.

It is a raw format, but not in the sense that you get the
direct output of each sensel.

It simply is not a raw format.

It is not an image, it has to be cooked to provide an image,
therefore it's raw.

It is interpolated data, not
raw data that needs to be interpolated.

Could you expand on that?


In practice those that use it,
the advantages they use RAW for, but without having to store
all the data they don't care for in first place.

It's just a TIFF RGB format. Whoopee. They could shoot
JPEG to get the same effect...

News for you: CR2 is a JPEG RGB format.


What's not clear to me, is whether that makes a sharper low res image in
the same way reducing after raw conversion will do.

Resizing to a smaller resolution does not make an image
"sharper" as such.

It thins the borders between ajdacent areas, just as wavelet
sharpen does. So not only the acutancy can rise.

No reply?


It removes high frequency detail,

which, depending on the image and the amount of downsizing,
isn't there in first place (excluding high frequency noise).

Well, its true that if you shoot pictures of gray cards, there
isn't much high frequency detail. Otherwise there is.

Please do some FT on real world images, and prove your claim.


From bayer patterns being very good but not perfect in the
restoration of high frequencies ( SQRT(2) of the pixel to pixel
distance) over the roll-off of AA filters down to camera shake,
slight misfocus and subject movement ... there can be many reasons
why high frequency detail simply isn't there in first place.

But for good photographers, who use good technique, there is
almost always enough high frequency detail to make a very
visible difference.

A "visible difference" between what and what? Do try a double
blind test on printed (or web sized) images, one 'original'
from bayer and one downsized to 70% ...

That is *precisely* why applying Sharpen almost always has a
significant effect.

'significant' meaning you can measure the difference, but you
cannot see it? 'significant' meaning a difference of several
orders of magnitude?

BTW, whether Sharpen has a 'significant effect' depends very much
on where the high pass filter cuts off lower frequencies.


and in
that sense reduces sharpness.

Which is irrelevant, unless you're actually seeing it at a
distance and resolution where such detail, if it was were,
would be visible. Most images just aren't viewed at 100%.

Prints commonly are.

'viewing distance'. 'loupe'.


Most photos aren't
all that sharp at full resolution, if for no other reason than the
existence of an antialiasing filter.

That isn't really true. The AA filter, in a properly processed
image, has just about exactly the same amount of high frequency
detail as a similar camera without the AA filter, except that at
frequencies very close to the Nyquist limit the signal to noise
ratio will be slightly reduced on the camera with the AA filter
(and conversely on camera without the AA filter the SNR will be
reduced by aliasing distortion throughout the frequency
spectrum).

That assumes a hard cut-off of the AA filter. I understand
that that's pretty hard to do in the real world.

It is virtually impossible to do, but that is not required
for what was described.

The high frequency components just below the Nyquist Limit are
reduced by the AA filter, but by no means eliminated because the
filter is not particularly sharp. Because of that the actual
frequency where maximum frequency distortion occurs will
intentionally be placed just above the Nyquist Limit. Most
designs will hit a minimum and at higher frequencies will have a
slope similar to the slope at lower frequencies. (On some that
slope may be very steep at frequencies well above the Nyquist
Limit, but hopefully that will also be at the upper limits for
any given lens that might be used, so between the filter and the
lens there is a very low response level at those frequencies
too.)

Whatever, the AA filter is a Low Pass filter, and a High Pass
filter can be applied in post processing with an almost opposite
frequency response to perfectly correct the rolloff provided
by the AA filter. At frequencies just below the Nyquist Limit (and
be aware that here are absolutely *no* frequencies above the Limit) the
correction from the Sharpen HP filter is at its highest. The filter
increases noise just as much as it increases desired signal, and
that is why the SNR is worse than it would without the AA filter but
the frequency response will be almost precisely the same.

Of course at lower and lower frequencies there is less and less
correction by the Sharpen filter, and therefore less and less
change to the SNR compared to an image taken without the AA filter.

So, where do I get a high pass filter that works exactly to
counteract the AA filter, and why do other filters provide
better results when the result is inspected visually?


In any case the reason images are apparently not sharp when
viewed at 100% is because a Bayer Color Filter encoded camera
simply cannot produce a tone transition in less than some number
of pixels, and the larger the demosiacing matrix the higher the
minimum for transition, as well as the more accurate the colors.

Actually, you are wrong here. Intelligent demosaicing
detects borders and doesn't smear them, irrespective of their
size. It's not a simple averaging process.

You simply cannot get a one pixel length tone transition,
without applying a Sharpen HP or USM filter. Yes different
demosaicing algorithms can produce sharper results, but absent
some form of sharpen, there are no truly "sharp" transitions.

What is 'truly "sharp"'? A single hot pixel? An aliased
line?

I see lots of transitions from A to B with just one pixel in
between --- and since I cannot align the camera pixel borders
perfectly with the image ...

Actually, looking at resolution tests, one gets transitions
of 1.5 and less pixels.


Of course that isn't really "sharpness", but acutance, and again
it can be increased with either a high pass filter (Sharpen) or
application of Unsharp Mask.

Of course that's really sharpness, as such an averaging process
would reduce or destroy the visibility of e.g. thin parallel
lines, thus reducing resolution drastically.

There is no actual increase in resolution. All that happens
with either USM or an HP filter is that the difference between
existing tone transition edges in increased.

Ah, how exactly do you define 'resolution' and how do you
define MTF?

-Wolfgang