«

Wolfgang Weisselberg <(E-Mail Removed)> wrote:
> Chris Malcolm <(E-Mail Removed)> wrote:
>> Wolfgang Weisselberg <(E-Mail Removed)> wrote:
>>> Chris Malcolm <(E-Mail Removed)> wrote:
>>>> Wolfgang Weisselberg <(E-Mail Removed)> wrote:

>>> Take any true-type (i.e. vector) font and look at it 33% smaller
>>> (if only by increasing the observation distance). Of course it'll
>>> be harder to read, even though no loss of detail has happened.

>> I avoid that effect by comparing them at the same viewing size, using
>> simple pixel magnification.

> "simple" magnification (e.g. nearest neighbour) will give
> suboptimal results.

That can be true in certain cases at certain magnifications. Easy
enough to check that and sidestep any misleading artefacts by simply
shifting up and down the scale.

>>> I also find myself often enlarging images (e.g. from other people)
>>> to 200% or 400% to better judge fine detail --- even though no
>>> more detail is being made visible.

>> Exactly. In order to be sure of what is going on I will often make the
>> pixels visible.

> Ans that can be self-defeating again. In most of my photographs
> I'm not concerned about visible pixels.
> Instead, I'd be rather concerned if pixels were visible. :-)

Of course. But to check out pixel level artefacts and find out where
they're coming from you have to look at the pixels. Just as although
you have no aesthetic interest photographs of test charts, they can
come in handy when checking out a lens.

> Only with that understanding can one undertake the trip to 100+%
> views and come out wiser --- and if some frame does not contain
> high frequency details, downsampling it so it only encodes the
> frequencies it contains does not cause appreciable data loss.

Exactly. That's why I use downsampling as a quick test of the
resolution achieved in a particular image, and why I use simple
downsampling amounts which correspond to simple vulgar fractions.

> (Remember that while interpolating from bayer sensors does give
> very good results, it cannot encode data it doesn't see without
> inventing it. Which is one reason why RAW (outside Foveon)
> doesn't save RGB triplets.).

Of course. And one reason for not saving RGB triplets from Bayer
arrays is that there is more detail resolution than colour resolution
in the array which any specific triplet translation method necessarily
compromises.

--
Chris Malcolm

Chris Malcolm <(E-Mail Removed)> wrote:
> Wolfgang Weisselberg <(E-Mail Removed)> wrote:
>> Chris Malcolm <(E-Mail Removed)> wrote:
>>> Wolfgang Weisselberg <(E-Mail Removed)> wrote:
>>>> Chris Malcolm <(E-Mail Removed)> wrote:
>>>>> Wolfgang Weisselberg <(E-Mail Removed)> wrote:

>>> I avoid that effect by comparing them at the same viewing size, using
>>> simple pixel magnification.

>> "simple" magnification (e.g. nearest neighbour) will give
>> suboptimal results.

> That can be true in certain cases at certain magnifications.

Even in all magnifications in certain cases.

> Easy
> enough to check that and sidestep any misleading artefacts by simply
> shifting up and down the scale.

Better to use a competent magnification in the first place.

And interesting to use some magnification method that invents
details, just to countercheck the assumption that you see true
detail in all cases.

>>> Exactly. In order to be sure of what is going on I will often make the
>>> pixels visible.

>> Ans that can be self-defeating again. In most of my photographs
>> I'm not concerned about visible pixels.
>> Instead, I'd be rather concerned if pixels were visible. :-)

> Of course. But to check out pixel level artefacts and find out where
> they're coming from you have to look at the pixels.

True.

> Just as although
> you have no aesthetic interest photographs of test charts, they can
> come in handy when checking out a lens.

Also true. Though I seem to remember at least some test charts
coming from "how much detail can one resolve in aerial photography
(i.e. what can one detect) (and what is just the effect of
film emulsion and humans 'seeing' things)" rather than aesthetic
interests. ;-)
In other words, pixel peeping is fine ... within limits.

> That's why I use downsampling as a quick test of the
> resolution achieved in a particular image, and why I use simple
> downsampling amounts which correspond to simple vulgar fractions.

I'd rather find out what frequencies are in the image
(fourier transform) rather than downsampling (which has to be
done competently, again). But that's me.

>> (Remember that while interpolating from bayer sensors does give
>> very good results, it cannot encode data it doesn't see without
>> inventing it. Which is one reason why RAW (outside Foveon)
>> doesn't save RGB triplets.).

> Of course. And one reason for not saving RGB triplets from Bayer
> arrays is that there is more detail resolution than colour resolution
> in the array

"Green" (which is actually not too insensitive for red and
blue) carries most resolution, and occupies 50% of the array.
It's distance to the next green pixel is SQRT(2), which is
ca. 1/70% ...

> which any specific triplet translation method necessarily
> compromises.

Of course you are wrong here. Just generate a TIFF from your RAW
with the same numbers of pixels, just store RGB triplets instead
of place-dependent luminosity-behind-$COLORFILTER-values.

Presto, a triplet translation which doesn't loose data (it's
justr wasteful) and can be turned into JPEGs, too.

-Wolfgang

Wolfgang Weisselberg <(E-Mail Removed)> wrote:
> Chris Malcolm <(E-Mail Removed)> wrote:
>> Wolfgang Weisselberg <(E-Mail Removed)> wrote:
>>> Chris Malcolm <(E-Mail Removed)> wrote:
>>>> Wolfgang Weisselberg <(E-Mail Removed)> wrote:
>>>>> Chris Malcolm <(E-Mail Removed)> wrote:
>>>>>> Wolfgang Weisselberg <(E-Mail Removed)> wrote:

>>>> I avoid that effect by comparing them at the same viewing size, using
>>>> simple pixel magnification.

>>> "simple" magnification (e.g. nearest neighbour) will give
>>> suboptimal results.

>> That can be true in certain cases at certain magnifications.

> Even in all magnifications in certain cases.

But not in cases where the magnification is simply of only a single
pixel, such as 4 times magnification, where one pixel becomes
represented by four of the same value. And so on. That's why I choose
that kind of very simple magnification for viewing magnification
changes.

>> Easy
>> enough to check that and sidestep any misleading artefacts by simply
>> shifting up and down the scale.

> Better to use a competent magnification in the first place.

Which is as I explained what I do. It never interpolates or blends
when the magnification factor does not demand it.

> And interesting to use some magnification method that invents
> details, just to countercheck the assumption that you see true
> detail in all cases.

I wouldn't find that interesting, which is why I don't do it.

>>>> Exactly. In order to be sure of what is going on I will often make the
>>>> pixels visible.

>>> Ans that can be self-defeating again. In most of my photographs

>> That's why I use downsampling as a quick test of the
>> resolution achieved in a particular image, and why I use simple
>> downsampling amounts which correspond to simple vulgar fractions.

> I'd rather find out what frequencies are in the image
> (fourier transform) rather than downsampling (which has to be
> done competently, again). But that's me.

Fourier transforms don't work for me because they can't tell the
difference between image detail and noise.

>>> (Remember that while interpolating from bayer sensors does give
>>> very good results, it cannot encode data it doesn't see without
>>> inventing it. Which is one reason why RAW (outside Foveon)
>>> doesn't save RGB triplets.).

>> Of course. And one reason for not saving RGB triplets from Bayer
>> arrays is that there is more detail resolution than colour resolution
>> in the array

> "Green" (which is actually not too insensitive for red and
> blue) carries most resolution, and occupies 50% of the array.
> It's distance to the next green pixel is SQRT(2), which is
> ca. 1/70% ...

I was talking about non-chromatic detail, which doesn't have to reach
as far as the next sensor of the same colour.

>> which any specific triplet translation method necessarily
>> compromises.

> Of course you are wrong here. Just generate a TIFF from your RAW
> with the same numbers of pixels, just store RGB triplets instead
> of place-dependent luminosity-behind-$COLORFILTER-values.

> Presto, a triplet translation which doesn't loose data (it's
> justr wasteful) and can be turned into JPEGs, too.

You might be right there. I've never played with TIFFs. I have noted
that different RAW converters give different levels of detail
resolution, in some special cases quite marked. And sometimes a RAW
converter will offer an update with improved translation for the RAW
from a specific camera, and the reviewers exclaim at the extra detail
it gets.

--
Chris Malcolm

Chris Malcolm <(E-Mail Removed)> wrote:
> Wolfgang Weisselberg <(E-Mail Removed)> wrote:
>> Chris Malcolm <(E-Mail Removed)> wrote:
>>> Wolfgang Weisselberg <(E-Mail Removed)> wrote:

>>>> "simple" magnification (e.g. nearest neighbour) will give
>>>> suboptimal results.

>>> That can be true in certain cases at certain magnifications.

>> Even in all magnifications in certain cases.

> But not in cases where the magnification is simply of only a single
> pixel, such as 4 times magnification, where one pixel becomes
> represented by four of the same value.

Nearest neighbour --- and that's what it is --- will give
suboptimal results for most viewing cases. This includes
upsampling for compatison with images with more pixels.

> And so on. That's why I choose
> that kind of very simple magnification for viewing magnification
> changes.

Take a smooth gradient between black and white. Downsample it,
upsample (or magnify) your way and you get banding. Upsample
properly and you get no banding.

>>> Easy
>>> enough to check that and sidestep any misleading artefacts by simply
>>> shifting up and down the scale.

>> Better to use a competent magnification in the first place.

> Which is as I explained what I do. It never interpolates or blends
> when the magnification factor does not demand it.

I respectfully disagree --- your choosen magnification is
incompetent for photography and twice incompetent for comparing
photographs of different pixel density.

>> And interesting to use some magnification method that invents
>> details, just to countercheck the assumption that you see true
>> detail in all cases.

> I wouldn't find that interesting, which is why I don't do it.

Of course you do, but the details your magnification invents
(banding, borders, hard change) are rather drab

>> I'd rather find out what frequencies are in the image
>> (fourier transform) rather than downsampling (which has to be
>> done competently, again). But that's me.

> Fourier transforms don't work for me because they can't tell the
> difference between image detail and noise.

That must be strong noise if it overpowers detail even in test
shots designed for maximum detail and minimum noise.

>> "Green" (which is actually not too insensitive for red and
>> blue) carries most resolution, and occupies 50% of the array.
>> It's distance to the next green pixel is SQRT(2), which is
>> ca. 1/70% ...

> I was talking about non-chromatic detail, which doesn't have to reach
> as far as the next sensor of the same colour.

And what happens to that detail once it hits the AA filter?

> I have noted
> that different RAW converters give different levels of detail
> resolution, in some special cases quite marked.

It depends a lot on how conservative they are with the input ---
if they don't mind guessing, they are going o have more detail
(and more *false* detail) in their results. Inventing details
can be a workable strategz for:

> And sometimes a RAW
> converter will offer an update with improved translation for the RAW
> from a specific camera, and the reviewers exclaim at the extra detail
> it gets.

such accolades.

-Wolfgang

Wolfgang Weisselberg <(E-Mail Removed)> wrote:
> Chris Malcolm <(E-Mail Removed)> wrote:
>> Wolfgang Weisselberg <(E-Mail Removed)> wrote:
>>> Chris Malcolm <(E-Mail Removed)> wrote:
>>>> Wolfgang Weisselberg <(E-Mail Removed)> wrote:

>>>>> "simple" magnification (e.g. nearest neighbour) will give
>>>>> suboptimal results.

>>>> That can be true in certain cases at certain magnifications.

>>> Even in all magnifications in certain cases.

>> But not in cases where the magnification is simply of only a single
>> pixel, such as 4 times magnification, where one pixel becomes
>> represented by four of the same value.

> Nearest neighbour --- and that's what it is --- will give
> suboptimal results for most viewing cases. This includes
> upsampling for compatison with images with more pixels.

>> And so on. That's why I choose
>> that kind of very simple magnification for viewing magnification
>> changes.

> Take a smooth gradient between black and white. Downsample it,
> upsample (or magnify) your way and you get banding. Upsample
> properly and you get no banding.

>>>> Easy
>>>> enough to check that and sidestep any misleading artefacts by simply
>>>> shifting up and down the scale.

>>> Better to use a competent magnification in the first place.

>> Which is as I explained what I do. It never interpolates or blends
>> when the magnification factor does not demand it.

> I respectfully disagree --- your choosen magnification is
> incompetent for photography and twice incompetent for comparing
> photographs of different pixel density.

>>> And interesting to use some magnification method that invents
>>> details, just to countercheck the assumption that you see true
>>> detail in all cases.

>> I wouldn't find that interesting, which is why I don't do it.

> Of course you do, but the details your magnification invents
> (banding, borders, hard change) are rather drab

I think there's a confusion going on between upsampling and
downsampling for purposes of viewing an image, and doing so in order
to change the size of the image in pixels. As I explained earlier I
use different methods for those two different purposes.

>>> I'd rather find out what frequencies are in the image
>>> (fourier transform) rather than downsampling (which has to be
>>> done competently, again). But that's me.

>> Fourier transforms don't work for me because they can't tell the
>> difference between image detail and noise.

> That must be strong noise if it overpowers detail even in test
> shots designed for maximum detail and minimum noise.

I thought you were more interested in real images than pixel peeping
on test charts taken in ideal conditions? In the real images I deal
with compromise between reducing noise and preserving detail
resolution is an everyday problem.

>>> "Green" (which is actually not too insensitive for red and
>>> blue) carries most resolution, and occupies 50% of the array.
>>> It's distance to the next green pixel is SQRT(2), which is
>>> ca. 1/70% ...

>> I was talking about non-chromatic detail, which doesn't have to reach
>> as far as the next sensor of the same colour.

> And what happens to that detail once it hits the AA filter?

You're asking theoretical questions and giving answers based on
simplifying mathematical assumptions when I presume you actually do
have a good DSLR and at least one lens with sufficiently good central
precision that critical cases can be found where downsizing the image
by 30% loses some central detail resolution? My impression is that
that there's a few such lenses in the lens armoury of most DSLR makes.

I'd be prepared to make an exception if your camera is the Canon EOS
7D :-)

--
Chris Malcolm

Chris Malcolm <(E-Mail Removed)> wrote:
> Wolfgang Weisselberg <(E-Mail Removed)> wrote:
>> Chris Malcolm <(E-Mail Removed)> wrote:
>>> Wolfgang Weisselberg <(E-Mail Removed)> wrote:
>>>> Chris Malcolm <(E-Mail Removed)> wrote:
>>>>> Wolfgang Weisselberg <(E-Mail Removed)> wrote:


>>>> And interesting to use some magnification method that invents
>>>> details, just to countercheck the assumption that you see true
>>>> detail in all cases.

>>> I wouldn't find that interesting, which is why I don't do it.

>> Of course you do, but the details your magnification invents
>> (banding, borders, hard change) are rather drab

> I think there's a confusion going on between upsampling and
> downsampling for purposes of viewing an image, and doing so in order
> to change the size of the image in pixels.

Explain to me the difference between up/downsampling an image to
view or print it at a certain number of pixels and up/downsampling
an image to a certain number of pixels (with the assumption that
the image *will* be viewed or printed at some time, maybe with
another up/downsampling step, as needed).

> As I explained earlier I
> use different methods for those two different purposes.

I don't see the difference, sorry

>>> Fourier transforms don't work for me because they can't tell the
>>> difference between image detail and noise.

>> That must be strong noise if it overpowers detail even in test
>> shots designed for maximum detail and minimum noise.

> I thought you were more interested in real images than pixel peeping
> on test charts taken in ideal conditions?

I am.

> In the real images I deal
> with compromise between reducing noise and preserving detail
> resolution is an everyday problem.

If the very best test shots in ideal circumstances only resolve X,
how do you expect to resolve more than X in the ordinary case?
And how can you then not recognize that if you only resolve X,
anything with higher frequencies must be noise, and hence can be
dropped without regret?

>>> I was talking about non-chromatic detail, which doesn't have to reach
>>> as far as the next sensor of the same colour.

>> And what happens to that detail once it hits the AA filter?

> You're asking theoretical questions

If you insist ... "Please provide a test shot in ideal conditions
with *your* gear, made to contain the highest frequencies you
can record with your camera, as noise free as you can make them.
Then run a FFT over that and kindly tell us what the highest
frequencies are that you can resolve. Thank you."

And kindly advise how that non-chromatic detail, which is spread
by a competent AA filter, doesn't reach as far as the next sensor
of the same colour ... and if it doesn't reach the next sensor,
how can it be displayed in correct colour? Assume that for all
relevant cases the green pixels can detect the non-chromatic
detail, and that their distance is 1.4 pixels from each other.

-Wolfgang

Wolfgang Weisselberg <(E-Mail Removed)> wrote:
> Chris Malcolm <(E-Mail Removed)> wrote:
>> Wolfgang Weisselberg <(E-Mail Removed)> wrote:
>>> Chris Malcolm <(E-Mail Removed)> wrote:
>>>> Wolfgang Weisselberg <(E-Mail Removed)> wrote:
>>>>> Chris Malcolm <(E-Mail Removed)> wrote:
>>>>>> Wolfgang Weisselberg <(E-Mail Removed)> wrote:

>>>>> And interesting to use some magnification method that invents
>>>>> details, just to countercheck the assumption that you see true
>>>>> detail in all cases.

>>>> I wouldn't find that interesting, which is why I don't do it.

>>> Of course you do, but the details your magnification invents
>>> (banding, borders, hard change) are rather drab

>> I think there's a confusion going on between upsampling and
>> downsampling for purposes of viewing an image, and doing so in order
>> to change the size of the image in pixels.

> Explain to me the difference between up/downsampling an image to
> view or print it at a certain number of pixels and up/downsampling
> an image to a certain number of pixels (with the assumption that
> the image *will* be viewed or printed at some time, maybe with
> another up/downsampling step, as needed).

This is where I came in, I think. I explained that a long time
ago. You've confused yourself by your use of the word "number of
pixels" in the above paragraph. You're making assumptions I don't
make. If by any chance you're really interested in my answer to the
above question, go back and read my original explanation.

--
Chris Malcolm

Chris Malcolm <(E-Mail Removed)> wrote:
> Wolfgang Weisselberg <(E-Mail Removed)> wrote:
>> Chris Malcolm <(E-Mail Removed)> wrote:

>>> I think there's a confusion going on between upsampling and
>>> downsampling for purposes of viewing an image, and doing so in order
>>> to change the size of the image in pixels.

>> Explain to me the difference between up/downsampling an image to
>> view or print it at a certain number of pixels and up/downsampling
>> an image to a certain number of pixels (with the assumption that
>> the image *will* be viewed or printed at some time, maybe with
>> another up/downsampling step, as needed).

> This is where I came in, I think.

Aeh, nope. You came in on fine details.

> I explained that a long time
> ago.

You explained that you prefer nearest neighbour for, ah,
better judging details by producing blocky upsizing artifacts.
You'll note that that is neither "for viewing an image" nor
"to change the size of the image in pixels".

> You've confused yourself by your use of the word "number of
> pixels" in the above paragraph.

So you change the size of an image in pixels, but don't change
the numbers of pixels?? Did *you*, by chance, mean 'change the
apparent pixel size of an image' (which of course increases the
number of pixels, usually by a power of 4)? If so, you confused
me by your completely muddled word choice, then pointed to me as
the culprit. Good job, really!

> You're making assumptions I don't
> make.

One of my assumptions was that you'd choose your words carefully
and intended to express what they convey. Should I not use that
assumption when reading what you write?

> If by any chance you're really interested in my answer to the
> above question, go back and read my original explanation.

If by any chance you're really interested in my answer to
your original explanation, go read my answer to your original
explanation. I understood you perfectly well, thank you very much,
but you seem intent on worst case upsampling for pixel peeping.
Of course that brings suboptimal results, like presumed resolution
loss, if you upsize by e.g. 141% ...

-Wolfgang