Resolution and pixel size not directly related?

Thread view:

Enable EMail Alerts

Start New Thread

Thread rating:

Rich - 14 Dec 2005 02:21 GMT

Pixel size should determine how much detail (resolution) an image
presents. A 12m sensor using a 100mm lens should out-resolve another
12m sensor with a larger pixel size using the same lens, the only
difference being the larger sensor would show a greater angle of view,
which in-turn means "fewer" pixels available for any given object in
the field of view..

But we (for whatever reason) don't see this directly translating to
proportional increases in resolution, at least according to most tests
done between different cameras.

For instance, a Nikon D2x should show greater resolution on a specific
target than a Canon 1DsMKII because although the Nikon has fewer
pixels (12m versus 16m or so) it will put "more" pixels on a given
object in the field of view of a 100mm lens. As a for instance, say
you were taking a picture of a person's face with the 100mm lens.
If you filled the frame with their face, you would have slightly less
than 12m pixels being used (lets say 10m) to comprise the face.

With the Canon, you'd have only about 7m pixels because although the
entire frame would show the face and a greater area around the face,
the face itself would be composed of fewer pixels because the pixels
in the Canon are larger than those in the Nikon.

So, how come we aren't seeing a direct correlation between resolution
and pixel pitches with given lenses? Point in fact, the Olympus E-500
"should" be providing the highest resolution of any DSLR simply
because it's pixels are smaller than those in most other DSLRs. We
know this does mean a greater propensity to produce noise at higher
ISOs than cameras with larger pixels, but why is resolution not
increased in proportion, at least according to the tests?

Reply to this Message

Bart van der Wolf - 14 Dec 2005 13:57 GMT

> Pixel size should determine how much detail (resolution) an image
> presents. A 12m sensor using a 100mm lens should out-resolve
[quoted text clipped - 4 lines]
> which in-turn means "fewer" pixels available for any given object in
> the field of view..

Since you seem to have at least thought about it, I'll react and
correct some of your misunderstandings/oversights before they lead to
the wrong conclusions (hence your being puzzled).

1. Sensels (Sensor Elements that make up the sensor array) have a
physical size which limits resolution. One cannot capture finer detail
than the size of the sensel, in a single capture.
2. Sensor array size will determine output magnification for a given
output size. Less magnification means better resolution in output.

So we have basically two relatively inverted contributors to
resolution.

There is little use in comparing resolution at different
Fields-of-View, because that is not relevant for photography. One
frames a scene for the best composition, so different sensor sizes
require different focal lenghts if one wants to retain perspective (by
shooting from the same position) and "fill the frame".

However, there is more to the image then what the lens projects on the
sensor array. The lens has aberrations, and suffers from physical
aperture defined diffraction.
That by itself also limits maximum resolution, and there is little
benefit to try and resolve detail with smaller sensors, if the
projection has no resolution to offer. What's more, there is also an
Anti-Aliasing filter in front of the sensor array, which spreads
detail to some degree over neighboring sensels.

The optimal resolution trade-off (between residual aberrations and
diffraction) for most DSLR lenses can be found between f/5.6 and f/11.
At e.g. f/8.0 the aperture induced refraction will restrict detail to
about a spotsize of 10.7 micron (it varies with wave-length). That
means that smaller sensels do not pick up more detail ! At best they
are able to sample the luminance distribution of the diffraction spot
a bit more accurately.

See for some more diffraction info the table at:
http://www.bobatkins.com/photography/technical/diffraction.html
Also see the Diffraction section (table 6) at:
http://clarkvision.com/imagedetail/does.pixel.size.matter/

> But we (for whatever reason) don't see this directly translating
> to proportional increases in resolution, at least according to
> most tests done between different cameras.

Besides the explanation above, the differences in AA-filter strength
and different post-processing (contrast and sharpening) will create
more confusion than clarity, when the test is poorly executed.
Resolution should be tested with MTF curves on Raw linearized gamma
data, and adjusted for output magnification if one wants to compare.

A concept like Linewidths per Picture Height (LPH), as used by
DPreview, does just that. It compensates for sensor array size
differences, in order to make sensible resolution comparisons
possible.

If one wants to compare Raw converters and post-processing, then the
test should be conducted with the same Raw image file, to be
meaningful. comparison of different camera Raws only makes sense after
resampling to identical magnification.

Given all the possible variables included in out-of-camera JPEGs, they
are a poor basis for resolution comparison, unless one wants to
compare JPEG camera quality (which has little to do with resolution
testing, but more with system testing).

Bart

Reply to this Message

Rich - 15 Dec 2005 01:09 GMT

>> Pixel size should determine how much detail (resolution) an image
>> presents. A 12m sensor using a 100mm lens should out-resolve
[quoted text clipped - 23 lines]
>require different focal lenghts if one wants to retain perspective (by
>shooting from the same position) and "fill the frame".

Ideally, but we know that isn't always possible. For instance, the
only nature photographers who can fill a frame with an animal
go to zoos!

>However, there is more to the image then what the lens projects on the
>sensor array. The lens has aberrations, and suffers from physical
[quoted text clipped - 4 lines]
>Anti-Aliasing filter in front of the sensor array, which spreads
>detail to some degree over neighboring sensels.

Yes, lens aberration is trade-off against aperture for resolution. But
I'm not sure numerically what the effect of the anti-aliasing filter
is. I remember someone saying Kodak DLSRs achieved high
resolution by not using a filter, but suffered from other optical
effects because it was left out.

>The optimal resolution trade-off (between residual aberrations and
>diffraction) for most DSLR lenses can be found between f/5.6 and f/11.
[quoted text clipped - 3 lines]
>are able to sample the luminance distribution of the diffraction spot
>a bit more accurately.

Would be interesting to know what the highest resolution lens (of a
given focal length) ever was, which lens controlled aberrations to the
point where all or most of the aperture could be used and thus produce
smaller spot sizes. The incorporation of newer glasses and surfaces
which deviate from spheres should allow the companies to produce
the best lenses ever, if they choose to.

>See for some more diffraction info the table at:
>http://www.bobatkins.com/photography/technical/diffraction.html
[quoted text clipped - 27 lines]
>
>Bart

Thanks for the explainations!
-Rich

Reply to this Message

Bart van der Wolf - 15 Dec 2005 03:51 GMT

SNIP

>>There is little use in comparing resolution at different
>>Fields-of-View, because that is not relevant for photography. One
[quoted text clipped - 6 lines]
> only nature photographers who can fill a frame with an animal
> go to zoos!

Not exclusively, but you would be amazed of the number of 'wild'-life
images taken at zoo's, like this 35m full-frame one;
<http://www.xs4all.nl/~bvdwolf/temp/CarimaCrisatata.jpg> taken at some
12 (!) inches with my 100mm Macro, through the wire netting.

SNIP

> Yes, lens aberration is trade-off against aperture for resolution.
> But
> I'm not sure numerically what the effect of the anti-aliasing filter
> is.

It varies by model. Canon models are generally (possible exception the
5D) considered to be on the 'soft' side due to adequate AA-filtering.
They typically allow significant post-processing to restore apparent
resolution.

> I remember someone saying Kodak DLSRs achieved high
> resolution by not using a filter, but suffered from other optical
> effects because it was left out.

Correct, the lack of an efficient AA-filter may introduce "impossible
to correct in post-processing" artifacts.

SNIP

> Would be interesting to know what the highest resolution lens
> (of a given focal length) ever was, which lens controlled
> aberrations to the point where all or most of the aperture could
> be used and thus produce smaller spot sizes.

Physical aperture will remain a limitation, due to diffraction, but it
in general is easier to tackle with 50+ mm fixed focal lenses.
Internal focus lenses could be usable at more than only their optimal
focus distance.

> The incorporation of newer glasses and surfaces which
> deviate from spheres should allow the companies to produce
> the best lenses ever, if they choose to.

Incorporation of aspherical and (synthetic) rare earth elements is
already being employed, but it usually is the combination of (various)
elements that makes flat-field and low aberration imaging possible.
Diffraction remains a physical limitation.

Bart

Reply to this Message

Roger N. Clark (change username to rnclark) - 15 Dec 2005 04:59 GMT

> Ideally, but we know that isn't always possible. For instance, the
> only nature photographers who can fill a frame with an animal
> go to zoos!

This is simply not true. For example, these are not in zoos,
are wild, and are full frame:

http://www.clarkvision.com/galleries/gallery.bear/web/brown_bear.c09.08.2004.JZ3
F2296.b-700.html

http://www.clarkvision.com/galleries/gallery.bear/web/brown_bear.c09.08.2004.JZ3
F2632.b-700.html

http://www.clarkvision.com/galleries/gallery.bird/web/c01.12.2003.img_4084.b-600.html

http://www.clarkvision.com/galleries/gallery.bird/web/sandhill.cranes.c12.01.200
4.JZ3F7332.arl.d-700.html

http://www.clarkvision.com/galleries/gallery.foxes/web/fox.c01.06.2004.JZ3F3731.
frmraw.2x.c-600.html

http://www.clarkvision.com/galleries/gallery.foxes/web/c11112002.IMG_0181b-600.html

http://www.clarkvision.com/galleries/gallery.mtn_goat/web/c071302.01.36-600.mtn.
goat.html

http://www.clarkvision.com/galleries/gallery.mtn_goat/web/c070899_04_20-mtngoat.html

And these are not unique. Many many many photographers do similar,
and even better work with wild animals in wild environments.

Roger

Reply to this Message

Rich - 16 Dec 2005 01:17 GMT

>> Ideally, but we know that isn't always possible. For instance, the
>> only nature photographers who can fill a frame with an animal
[quoted text clipped - 23 lines]
>
>Roger

Sorry, I was overstating the case, but pictures I've often seen in
National Geographic, etc, show clear signs of heavy cropping.
Nice pictures, especially the Kodiak eating the salmon.
-Rich

Reply to this Message

Roger N. Clark (change username to rnclark) - 15 Dec 2005 04:47 GMT

> A concept like Linewidths per Picture Height (LPH), as used by DPreview,
> does just that. It compensates for sensor array size differences, in
> order to make sensible resolution comparisons possible.

Bart, I don't believe the LPH values on dpreview always have
meaning in reality. The problem is that the in camera processing
enhances MTF response to high contrast test targets giving
a false sense of real resolution. Some of the LPH is almost
at the theoretical maximum assuming the pixels were aligned perfectly
with the lines in the test target and there wass no anti-alias filter, and
optics were perfect. It is just too good to be true, and in fact
is when compared to resolution of more complex detail in real images.

This effect can be seen on my lens testing page:
http://www.clarkvision.com/imagedetail/relative-lens-sharpness

Note the high contrast bars (either color or black and white), versus
the mixed color target on the right. The mixed color target does
poor in comparison to the bars on white.

Roger

Reply to this Message

Rich - 16 Dec 2005 01:14 GMT

>> A concept like Linewidths per Picture Height (LPH), as used by DPreview,
>> does just that. It compensates for sensor array size differences, in
[quoted text clipped - 17 lines]
>
>Roger

I remember reading Sidgwick's book about how resolution of
black on white backgrounds, and similar high contrast targets
can sometimes go 10x finer than is recommended by theory.
-Rich

Reply to this Message

Bart van der Wolf - 16 Dec 2005 01:20 GMT

>> A concept like Linewidths per Picture Height (LPH), as used by
>> DPreview, does just that. It compensates for sensor array size
[quoted text clipped - 3 lines]
> Bart, I don't believe the LPH values on dpreview always have
> meaning in reality.

Ah true, but that's a different subject. The LPH principle will
compensate for sensor array size differences, but the DPreview values
are based on bi-tonal patterns, more or less aligned with the sensor
array.

As the relevant ISO standard states, those hyperbolic patterns are
intended for visual comparisons, but not for quantitative evaluation.
The slanted rectangle edges are the ones from which the MTF can be
calculated.

> The problem is that the in camera processing enhances MTF response
> to high contrast test targets giving a false sense of real
> resolution.

Yes, that as well, but true resolution measurements need to be taken
from linearized gamma, unsharpened sensor data. Determination of the
so-called OECF, Opto-Electronic-Conversion-Function, should be part of
the official resolution test.

> Some of the LPH is almost at the theoretical maximum assuming the
> pixels were aligned perfectly with the lines in the test target and
> there wass no anti-alias filter, and
> optics were perfect. It is just too good to be true, and in fact
> is when compared to resolution of more complex detail in real
> images.

Yes, accidental (!) alignment of the pattern with the array will push
the results towards the theoretical limit, and there is no robust way
of separating signal from aliasing artifacts. The bi-tonal patterns
are not suited for a discrete sampling system.

> This effect can be seen on my lens testing page:
> http://www.clarkvision.com/imagedetail/relative-lens-sharpness
>
> Note the high contrast bars (either color or black and white),
> versus the mixed color target on the right. The mixed color target
> does poor in comparison to the bars on white.

Yes, you'll witness effects of (mis-)alignment with the Bayer CFA.
Also the diagonally more dense sampling of Green will skew results
that are green and/or luminance related.
Another side-effect is that even MTF determinations need to be
corrected if not taken from sinusoidal patterns
(http://www.normankoren.com/Tutorials/MTF5.html#MTF_4_pi).

Bart

Reply to this Message