Sharpening artefacts and MTF of monitors?

I'm trying to understand visible effects of artefacts of sharpening with
different viewing conditions. Suppose I have a sharpening matrix such
that I can't see artefacts on monitor (with a very good CRT monitor,
VS P815). However, when I blow pixels 200%, a careful examination
exposes some artefacts. With magnification about 500% artefacts become
very visible.

Given this data, can I estimate how the image will be seen on LCD monitors
(which have better MTF at high spacial frequencies)? Can I estimate
on which pixel-per-inch value the printed image will have visible
artefacts?

In short: what makes the artefact invisible without magnification: bad
MTF of the CRT monitor (due to internal reflection in glass), or bad MTF
of the eye? If the latter, then a print with about 120 pixels/inch or
more should show no artefacts (pixel size on the monitor is 0.26mm).
If the former, then one needs much higher pixels/inch value...

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

Here is an example: take the image

http://www.dpreview.com/reviews/Koni...res/ACRraw.jpg

This is a resolution chart, the artefacts should be very visible. I run
DSP software, and it shows the following MTF curve of the combined
lens/sensor/demosaicer:

1 """"xxx__''''''''''''''''''''''''''''''''''''''''' '''''''''''''|
| ""x_ |
| "x_ |
| "x_ |
| "x_ |
| "_ |
| "x_ |
| x_ |
| "x |
| "x_ |
| "_ |
| "x_ |
| "_ | |
| "x | |
| "x_ | |
| x_ | |
| "_ | |
| "x |
| |"x_ |
| | x_ |
| | "x_ |
-0.016------------------------------------------------------------"xx_
0 3000

(vertical line is the Nyquist frequency of the sensor). One can easily
see that sharpening with the matrix

0 -0.375 0
-0.375 2.5 -0.375
0 -0.375 0

makes the throughput MTF curve almost horizontal:


1 """"""""""""""""xxxx____'''''''''''''''''''''''''' '''''''''''''|
| ""xx__ |
| ""xx_ |
| "xx_ |
| "_ |
| "x |
| "x |
| "_ |
| x |
| "_ | |
| _| |
| x |
| |x |
| | x |
| | x |
| | " |
| | " |
| | " |
| | "_ |
| | _ |
,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,, ,|,,,,,,,,,x,,
-.05|.............................................. ....|.........."_
0 3000

Indeed, this sharpening has enormous "clearing" effect on the picture;
the artefacts are as described above: invisible on CRT without
magnification (initial image has no artefacts). It would be
interesting to know what happens with other media...

Thanks,
Ilya

In article ,
Ilya Zakharevich wrote:

I'm trying to understand visible effects of artefacts of sharpening with
different viewing conditions. Suppose I have a sharpening matrix such
that I can't see artefacts on monitor (with a very good CRT monitor,
VS P815). However, when I blow pixels 200%, a careful examination
exposes some artefacts. With magnification about 500% artefacts become
very visible.

Given this data, can I estimate how the image will be seen on LCD monitors
(which have better MTF at high spacial frequencies)? Can I estimate
on which pixel-per-inch value the printed image will have visible
artefacts?

In short: what makes the artefact invisible without magnification: bad
MTF of the CRT monitor (due to internal reflection in glass), or bad MTF
of the eye? If the latter, then a print with about 120 pixels/inch or
more should show no artefacts (pixel size on the monitor is 0.26mm).
If the former, then one needs much higher pixels/inch value...

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

Here is an example: take the image

http://www.dpreview.com/reviews/Koni...res/ACRraw.jpg

This is a resolution chart, the artefacts should be very visible. I run
DSP software, and it shows the following MTF curve of the combined
lens/sensor/demosaicer:

1 """"xxx__''''''''''''''''''''''''''''''''''''''''' '''''''''''''|
| ""x_ |
| "x_ |
| "x_ |
| "x_ |
| "_ |
| "x_ |
| x_ |
| "x |
| "x_ |
| "_ |
| "x_ |
| "_ | |
| "x | |
| "x_ | |
| x_ | |
| "_ | |
| "x |
| |"x_ |
| | x_ |
| | "x_ |
-0.016------------------------------------------------------------"xx_
0 3000

(vertical line is the Nyquist frequency of the sensor). One can easily
see that sharpening with the matrix

0 -0.375 0
-0.375 2.5 -0.375
0 -0.375 0

makes the throughput MTF curve almost horizontal:


1 """"""""""""""""xxxx____'''''''''''''''''''''''''' '''''''''''''|
| ""xx__ |
| ""xx_ |
| "xx_ |
| "_ |
| "x |
| "x |
| "_ |
| x |
| "_ | |
| _| |
| x |
| |x |
| | x |
| | x |
| | " |
| | " |
| | " |
| | "_ |
| | _ |
,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,, ,|,,,,,,,,,x,,
-.05|.............................................. ....|.........."_
0 3000

Indeed, this sharpening has enormous "clearing" effect on the picture;
the artefacts are as described above: invisible on CRT without
magnification (initial image has no artefacts). It would be
interesting to know what happens with other media...

Thanks,
Ilya




Sharpening could work in this case. In some cases you could see ripples
in that chart or the S/N ratio will get bad.

[A complimentary Cc of this posting was sent to
Kevin McMurtrie
], who wrote in article :

1 """"""""""""""""xxxx____'''''''''''''''''''''''''' '''''''''''''|
| ""xx__ |
| ""xx_ |
| "xx_ |
| "_ |
| "x |
| "x |
| "_ |
| x |
| "_ | |
| _| |
| x |
| |x |
| | x |
| | x |
| | " |
| | " |
| | " |
| | "_ |
| | _ |
,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,, ,|,,,,,,,,,x,,
-.05|.............................................. ....|.........."_
0 3000

Sharpening could work in this case. In some cases you could see ripples
in that chart or the S/N ratio will get bad.

Of course, S/N is decreased about 2.6 times by this sharpening. So it
depends on what is the initial value of S/N. I did not measure S/N of
the Adobe demosaicer; with the builtin demozaicer of the camera it is
about 60 (for luma of 18% gray, at ISO50). Probably such a good RTF
curve should result in larger noise than the lousy MTF curve of the
builtin demozaicer... Still, it may happen that S/N is above 20
(i.e., tolerable). Anyway, it is going to be a usual tradeoff between
detail and noise; note that one can process different parts of the
image with different settings.

About ripples: with MTF curve such as above, I doubt that any artefact
like "ripples" (whatever it is ;-) may be possible; there is a light
halo about dark areas (visible only under ). However, I was
seriously surprised about the quality of the image; having MTF above
50% corner-to-corner for an equivalent of 38 lp/mm (for 35mm film)
without visible artefacts gives very counterintuitive results...

Yours,
Ilya

I wrote in article :

Throughput MTF curve of lens+sensor+demosaicer+sharpening is

1 """"""""""""""""xxxx____'''''''''''''''''''''''''' '''''''''''''|
| ""xx__ |
| ""xx_ |
| "xx_ |
| "_ |
| "x |
| "x |
| "_ |
| x |
| "_ | |
| _| |
| x |
| |x |
| | x |
| | x |
| | " |
| | " |
| | " |
| | "_ |
| | _ |
,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,, ,|,,,,,,,,,x,,
-.05|.............................................. ....|.........."_
0 3000

Indeed, this sharpening has enormous "clearing" effect on the picture;
the artefacts are as described above: invisible on CRT without
magnification (initial image has no artefacts). It would be
interesting to know what happens with other media...

BTW, I temporarily put the sharpened picture on
http://ilyaz.org/software/tmp/KM_A20...sharpening.jpg

It is 1.3M file (saved with "85% quality"; should have I chosen a
better quality?).

Yours,
Ilya

P.S. I calculated MTF to the cut-off frequency of the lens (f/4):

1 """"""xx_''''''''''''''''''''''''''''''''''''''''' '''''''''''''|
| "x_ |
| "_ |
| x |
| x |
| x |
| _ | |
| | |
| x| |
| | |
| x |
| : |
| |: |
| |x |
| |: |
| | : |
| | " |
| | |
| | " __ _
0 -----------------------x-----__xxxxxxxxxxx_-----------x"--"xxx"-
| | "__x" "x_ _" |
-.123|............................................. "x___"...........|
0 8000

(the vertical line is at Nyquist of the sensor, anything to the right
results in aliasing; but the data too far to the right is probably not
very reliable; I doubt that the second "dip" in the graph is actually
there). It is increadibly good; very little aliasing and very high
MTF in the main area; anyone having an idea how Adobe does it?

[A complimentary Cc of this posting was sent to
Kevin McMurtrie
], who wrote in article :

Sharpening could work in this case. In some cases you could see ripples
in that chart or the S/N ratio will get bad.

I measured the S/N ratio after this sharpening. It looks like S/N of
18% gray should be 36. This is the S/N ratio of luma; thus S/N ratio
of luminosity is 16; so it coincides with S/N ratio of Velvia 50 (as
given on Roger Clark's web site).

Yours,
Ilya

[A complimentary Cc of this posting was sent to
Kevin McMurtrie
], who wrote in article :

Sharpening could work in this case. In some cases you could see ripples
in that chart or the S/N ratio will get bad.

I measured the S/N ratio after this sharpening. It looks like S/N of
18% gray should be 36. This is the S/N ratio of luma; thus S/N ratio
of luminosity is 16; so it coincides with S/N ratio of Velvia 50 (as
given on Roger Clark's web site).

Yours,
Ilya

Ilya Zakharevich wrote:
[A complimentary Cc of this posting was sent to
Kevin McMurtrie
], who wrote in article
:

Sharpening could work in this case. In some cases you could see
ripples in that chart or the S/N ratio will get bad.

I measured the S/N ratio after this sharpening. It looks like S/N of
18% gray should be 36. This is the S/N ratio of luma; thus S/N ratio
of luminosity is 16; so it coincides with S/N ratio of Velvia 50 (as
given on Roger Clark's web site).

Yours,
Ilya

But the broadband signal-to-noise ratio may not reflect how the image is
perceived. You need to know the narrowband SNR, and weight that with the
visual systems frequency response (as is done with audio and weighted SNR
measurements). This will be particularly important in comparing different
media (e.g. film and digital).

David

Ilya Zakharevich wrote:
[A complimentary Cc of this posting was sent to
Kevin McMurtrie
], who wrote in article
:

Sharpening could work in this case. In some cases you could see
ripples in that chart or the S/N ratio will get bad.

I measured the S/N ratio after this sharpening. It looks like S/N of
18% gray should be 36. This is the S/N ratio of luma; thus S/N ratio
of luminosity is 16; so it coincides with S/N ratio of Velvia 50 (as
given on Roger Clark's web site).

Yours,
Ilya

But the broadband signal-to-noise ratio may not reflect how the image is
perceived. You need to know the narrowband SNR, and weight that with the
visual systems frequency response (as is done with audio and weighted SNR
measurements). This will be particularly important in comparing different
media (e.g. film and digital).

David

[A complimentary Cc of this posting was sent to
David J Taylor
], who wrote in article :
I measured the S/N ratio after this sharpening. It looks like S/N of
18% gray should be 36. This is the S/N ratio of luma; thus S/N ratio
of luminosity is 16; so it coincides with S/N ratio of Velvia 50 (as
given on Roger Clark's web site).

But the broadband signal-to-noise ratio may not reflect how the image is
perceived. You need to know the narrowband SNR, and weight that with the
visual systems frequency response (as is done with audio and weighted SNR
measurements).

Thanks, this may be relevant in some other situation. But given that
MRF of the whole workflow is provided in the initial message, and is
"almost horizontal", weighting will not change things much on the
digital side.

On the film side - I know no data about narrow band SNR of Velvia 50;
do you? And the MTF curve of Velvia 50 + reasonably good fixed focal
length lens is going to be "similarly good" when you compensate for 4x
difference in sensor size, so the broadband S/N should be viable
too...

Anyway, it may be that your suggestions will change the numbers about
25%; but I expect the change to be in the same direction for digital
and film. Given that I suspect very much[*] the data for film noise
on Roger's site, such a correction is not very important.

Thanks,
Ilya
[*] E.g., his S/N=16 for 18% gray on Velvia 50

http://clarkvision.com/imagedetail/d...gnal.to.noise/

does not specify the window size into the film; he compares with a
camera which has 8M pixels, and 8.2 micron cells, so there may be
different interpretations of "equivalent" window size into the film.
In a private communication he says that the window size is actually
6.3 microns (equivalent of 24M pixels).

Actually, comparing noise of 24MP scan with noise of 8MP digital
sensor may be not that crazy, given his "translation rules" from scan
resolution to digital sensor resolution (in other papers on his site).
However, it is clear that one can reduce noise of 24MP scan by
post-processing without lowering the resolution much (comparing to 8MP
scan); I do not know whether this is taken into account.

Additionally, it is not specified whether S/N=16 is for noise of
density, or noise of "initial luminance". Given high contrast of
slide film (gamma = 1/1.5), the latter should be about 66% of the
former... Roger still did not answer my email about this issue.

[A complimentary Cc of this posting was sent to
David J Taylor
], who wrote in article :
I measured the S/N ratio after this sharpening. It looks like S/N of
18% gray should be 36. This is the S/N ratio of luma; thus S/N ratio
of luminosity is 16; so it coincides with S/N ratio of Velvia 50 (as
given on Roger Clark's web site).

But the broadband signal-to-noise ratio may not reflect how the image is
perceived. You need to know the narrowband SNR, and weight that with the
visual systems frequency response (as is done with audio and weighted SNR
measurements).

Thanks, this may be relevant in some other situation. But given that
MRF of the whole workflow is provided in the initial message, and is
"almost horizontal", weighting will not change things much on the
digital side.

On the film side - I know no data about narrow band SNR of Velvia 50;
do you? And the MTF curve of Velvia 50 + reasonably good fixed focal
length lens is going to be "similarly good" when you compensate for 4x
difference in sensor size, so the broadband S/N should be viable
too...

Anyway, it may be that your suggestions will change the numbers about
25%; but I expect the change to be in the same direction for digital
and film. Given that I suspect very much[*] the data for film noise
on Roger's site, such a correction is not very important.

Thanks,
Ilya
[*] E.g., his S/N=16 for 18% gray on Velvia 50

http://clarkvision.com/imagedetail/d...gnal.to.noise/

does not specify the window size into the film; he compares with a
camera which has 8M pixels, and 8.2 micron cells, so there may be
different interpretations of "equivalent" window size into the film.
In a private communication he says that the window size is actually
6.3 microns (equivalent of 24M pixels).

Actually, comparing noise of 24MP scan with noise of 8MP digital
sensor may be not that crazy, given his "translation rules" from scan
resolution to digital sensor resolution (in other papers on his site).
However, it is clear that one can reduce noise of 24MP scan by
post-processing without lowering the resolution much (comparing to 8MP
scan); I do not know whether this is taken into account.

Additionally, it is not specified whether S/N=16 is for noise of
density, or noise of "initial luminance". Given high contrast of
slide film (gamma = 1/1.5), the latter should be about 66% of the
former... Roger still did not answer my email about this issue.