Ad
Ad
Ad
Pages: « 1 [2]   Bottom of Page
Print
Author Topic: Everything matters, science, etc.  (Read 5177 times)
bdosserman
Full Member
***
Offline Offline

Posts: 120


« Reply #20 on: February 16, 2012, 11:23:15 AM »
ReplyReply

Hi,

I am certain that there is more to print sharpness/detail than the number of dpi. If you really want to establish good lower bounds based on empiry, you probably dont want to use a general purpose image processing pipeline that includes god knows how many image scalings, an unknown dithering pattern, and unknown ink splatter characteristics. Not that scaling in a non-linear color/intensity representation (such as sRGB) actually gives some (usually small) errors compared to moving the image closer/further away from the viewer.

This is a good point. To me, from a naive point of view this supports the argument that resolution increases beyond what the eye is normally capable of seeing could end up producing perceptible improvements to the final image.

Quote
I think that image perception research have suggested that 1 minute of arc is a good rule of thumb for a normal-good pair of eyes. I presume this is for relatively high-contrast black/white transitions perfectly smoothly/sharp printed on paper. This could be compared to actual prints at a certain size/distance/resolution, and if they are in the same ballpark, I would be satisfied.

Is this for what the eyes can "resolve", or the point at which they can no longer tell the difference between higher resolutions? I could imagine these could be pretty different. [Also, a side comment on "normal eyes", following up on what someone brought up in a different thread: my vision is terrible, but with glasses it corrects to about 20/15. I guess this is a noticeable difference from 20/20]

Quote
Sure. But when do you stop? Is it worthwhile to have an image that prints 1x1 meter, where you can step closer until you reach the close-focus limit of your eyesight and still see no resolution limitations?

No, obviously not. But I'm just pointing out that one should include some leeway for the presumed viewing distance.

Quote
"Megapixels" is a horrible concept. But I do believe that Bayer sensors behave very good for the kind of scenes that photographers tend to shoot, and for viewers that are human. Certainly a lot better than what a simplistic analysis of the red/green/blur sensels would suggest.

This certainly seems to be the case.

Quote
Does it have to be either or? I think that doing both and relating them to each other is a good approach.

Of course not. I find the theoretical discussion interesting, and ideally would like to have it and the empirical discussion developed to the point where both are in agreement. It just bothers me when people present a theoretical analysis as if it should close all further discussion. In general, but especially here where the estimates of what resolutions the eye can readily perceive are varying by factors of four, right in the range that completely changes the answer to the original question.

Quote
Oversampling is a method where you can replace a complex analog filter with a (equally complex but simpler to design) digital filter. Or you might see it (with dithering) as a method to provide many levels at lower sample-rates, when the physical technology is better suited for fewer levels and higher sample-rates.Scaling your typical over-sampled image is not that hard. Scaling an image that has been excessively over-sharpened/aliased through non-Nyquistian sampling is harder.

Can you explain the last sentence? I'm not familiar with the terminology.

Quote
Many image processing algorithms and printer drivers seems to use very naiive image scaling algorithms. Perhaps because most of their customers dont care that much.

It would be interesting to know more about this part of the pipeline, and how much it's affecting the final output quality.

Quote
So do you want blurry pixels that are a good approximation to the continous waveform, or sharp (aliased) pixels that have the "right value, but the wrong location"? I think that actually photography lags behind in this discussion, designers of computer fonts seems to be a lot more conscious of the trade-offs, perhaps because theoir "product" consists of 8x8 pixel "images" where one really can afford to obsess over single-pixel details.
http://www.joelonsoftware.com/items/2007/06/12.html

I think that's exactly why font designers obsess over it (that plus the obviously related point that their individual pixels are a whole lot more visible than those in a photographic print). I agree that these things ought to be considered more seriously (at least, assuming they haven't been; I don't claim to know what people have and haven't thought about), but with the obvious caveat that what makes sense for fonts is probably not what will make sense for photography, given the very different circumstances.

Brian
Logged
hjulenissen
Sr. Member
****
Offline Offline

Posts: 1678


« Reply #21 on: February 16, 2012, 04:34:26 PM »
ReplyReply

This is a good point. To me, from a naive point of view this supports the argument that resolution increases beyond what the eye is normally capable of seeing could end up producing perceptible improvements to the final image.
Perhaps. I would rather have reasonably firm limits on what we are capable of, and what the limitations of our tools are.
Quote
Is this for what the eyes can "resolve", or the point at which they can no longer tell the difference between higher resolutions? I could imagine these could be pretty different. [Also, a side comment on "normal eyes", following up on what someone brought up in a different thread: my vision is terrible, but with glasses it corrects to about 20/15. I guess this is a noticeable difference from 20/20]
http://en.wikipedia.org/wiki/Minute_of_arc#Human_vision
"In humans, 20/20 vision is the ability to resolve a spatial pattern separated by a visual angle of one minute of arc. A 20/20 letter subtends 5 minutes of arc total.
For raster graphics, Apple Inc asserts that a display of approximately 300 ppi at a distance of 12 inches (305 mm) from one's eye, or 57 arcseconds per pixel[9] is the maximum amount of detail that the human retina can perceive.[10] Raymond Soneira, president of DisplayMate Technologies, however, stated that the resolution of the human retina is higher than claimed by Apple, working out to 477 ppi at 12 inches (305 mm) or 36 arcseconds per pixel."
Quote
No, obviously not. But I'm just pointing out that one should include some leeway for the presumed viewing distance.
Sure. Or you might want to establish the most likely limits, then multiply that number by 1.001x, 2x or 4x, depending on what is deemed cost-efficient.
Quote
It just bothers me when people present a theoretical analysis as if it should close all further discussion.
Those with the firmest conviction may often tend not to be those with the most knowledge :-) I don't see what you are describing a lot in here.
Quote
Can you explain the last sentence? I'm not familiar with the terminology.
Sorry. Read "cameras lacking OLPF/AA", or "printer (drivers) using bilinear interpolation".
Quote
I think that's exactly why font designers obsess over it (that plus the obviously related point that their individual pixels are a whole lot more visible than those in a photographic print). I agree that these things ought to be considered more seriously (at least, assuming they haven't been; I don't claim to know what people have and haven't thought about), but with the obvious caveat that what makes sense for fonts is probably not what will make sense for photography, given the very different circumstances.
Circumstances may be different, but out vision is not. Therefore I think we might learn something from them.

Text fonts are very well known - we "know how they should look". Compared to the leafs of a tree that not even the photographer remembers exactly how they were laid out.

Also, the guy spending days designing the 'e' letter will have his 8x8 pixel efforts duplicated into billions of readers eyes. Hardly something that any 8x8 pixel region of my pictures can expect...

-h
Logged
bdosserman
Full Member
***
Offline Offline

Posts: 120


« Reply #22 on: February 16, 2012, 10:58:27 PM »
ReplyReply

Perhaps. I would rather have reasonably firm limits on what we are capable of, and what the limitations of our tools are.

Of course this would be nice to know. But it would be best to have really reliable limits. I guess all I'm saying is that if there's cumulative error introduced at each of several steps in the process, this should be understood and taken into account in these sorts of estimates. Obviously, it would be even better to minimize this error wherever possible.

Quote
http://en.wikipedia.org/wiki/Minute_of_arc#Human_vision
"In humans, 20/20 vision is the ability to resolve a spatial pattern separated by a visual angle of one minute of arc. A 20/20 letter subtends 5 minutes of arc total.
For raster graphics, Apple Inc asserts that a display of approximately 300 ppi at a distance of 12 inches (305 mm) from one's eye, or 57 arcseconds per pixel[9] is the maximum amount of detail that the human retina can perceive.[10] Raymond Soneira, president of DisplayMate Technologies, however, stated that the resolution of the human retina is higher than claimed by Apple, working out to 477 ppi at 12 inches (305 mm) or 36 arcseconds per pixel."

OK, so if I understand correctly this is pointing to precisely the issue I brought up -- there's potentially a substantial difference between where the eye can no longer distinguish individual pixels, and where it can no longer perceive the presence of additional resolution. Again, it would be nice if these things were clearly stating whether they were testing the eye's ability to resolve a given resolution, or the eye's ability to perceive the difference in resolution. (The latter would be a much simpler test: show a bunch of people images at two different resolutions, ask them to pick the one that looks smoother, and see where the accuracy drops to 50/50).

Quote
Sorry. Read "cameras lacking OLPF/AA", or "printer (drivers) using bilinear interpolation".

I'd like to understand better what you mean by "non-Nyquistian sampling". Is it a question of sampling at less than double the frequency of the original data?  Or I suppose in your context, not the original data, but the post-filter data. I don't understand the relevance of printer drivers to sampling.

In any case, it seems to me that downscaling by an integer multiple will be fine in this situation, which is the typical case for oversampling (not that I'm claiming this would have the same effect as the filter, but that's not where the topic came from). But obviously in practice one might need to downscale by something else, and I could image that causing more issues.

Quote
Circumstances may be different, but out vision is not. Therefore I think we might learn something from them.

It could be worthwhile, certainly. All I meant was that the typical font is viewed at a much lower resolution than the typical photo (and is obviously composed quite differently), so the aspects of human vision most relevant to the font designers could be quite different from those most relevant to photographers.

Brian
Logged
hjulenissen
Sr. Member
****
Offline Offline

Posts: 1678


« Reply #23 on: February 17, 2012, 12:10:31 AM »
ReplyReply

OK, so if I understand correctly this is pointing to precisely the issue I brought up -- there's potentially a substantial difference between where the eye can no longer distinguish individual pixels, and where it can no longer perceive the presence of additional resolution. Again, it would be nice if these things were clearly stating whether they were testing the eye's ability to resolve a given resolution, or the eye's ability to perceive the difference in resolution. (The latter would be a much simpler test: show a bunch of people images at two different resolutions, ask them to pick the one that looks smoother, and see where the accuracy drops to 50/50).
http://downloads.bbc.co.uk/rd/pubs/whp/whp-pdf-files/WHP092.pdf
"Tests of visual acuity to determine the resolution required of a television transmission system

The most important result is that, when using the sharp-cut “Rec. 601” filter, the average visible limit of edge rise-time lies in the range 0.997 to 1.111 minutes of arc, with a confidence of 95%. Interestingly, the generally accepted figure for visual acuity is approximately 1 minute of arc."

They used HD crt-screens at a sufficiently large distance where individual pixels cannot be seen, then experimented with different filters in still-images to see what kind of resolution-loss was indistinguishable from a side-by-side high-resolution reference.

Relevant to our stuff is that they find that sharpening can (to some degree) compensate for real resolution.
I'd like to understand better what you mean by "non-Nyquistian sampling". Is it a question of sampling at less than double the frequency of the original data?  Or I suppose in your context, not the original data, but the post-filter data. I don't understand the relevance of printer drivers to sampling.
If your data contains a bandwidth >= fs/2, then you need to reduce that bandwidth in order to do a sampling according to Nyquist. In text-books this is carried out with a perfect sinc brickwall lowpass filter. If you dont do this, you will have aliasing.

Any scaling is in practice a resampling process. If the printer (driver) receive a 24 MP image and the nozzle works at a corresponding grid that is 23MP, then the image needs to be resampled. This resampling needs to trade off passband (MTF50) vs stop-band (aliasing) while keeping an eye on spatial response (ringing).

The final ink pattern can be seen as the dual of the A/D conversion carried out in the camera sensor. The same theory that suggests OLPF in cameras, also suggest that the ink pattern on paper should be even and smooth - unless the resolution is high enough that e.g. our eyes can do the smoothing.

My printer views are theoretical. I have seen input and output, but have no other insight into the black box than reasoning.
Quote
It could be worthwhile, certainly. All I meant was that the typical font is viewed at a much lower resolution than the typical photo (and is obviously composed quite differently), so the aspects of human vision most relevant to the font designers could be quite different from those most relevant to photographers.
A valid point. But if we are exploring the limits of how close viewers may possibly get to the large print to investigate detail, may we not get into similar projected pixel angles?

-h
« Last Edit: February 17, 2012, 12:20:48 AM by hjulenissen » Logged
bdosserman
Full Member
***
Offline Offline

Posts: 120


« Reply #24 on: February 17, 2012, 09:19:19 PM »
ReplyReply

http://downloads.bbc.co.uk/rd/pubs/whp/whp-pdf-files/WHP092.pdf
"Tests of visual acuity to determine the resolution required of a television transmission system

The most important result is that, when using the sharp-cut “Rec. 601” filter, the average visible limit of edge rise-time lies in the range 0.997 to 1.111 minutes of arc, with a confidence of 95%. Interestingly, the generally accepted figure for visual acuity is approximately 1 minute of arc."

They used HD crt-screens at a sufficiently large distance where individual pixels cannot be seen, then experimented with different filters in still-images to see what kind of resolution-loss was indistinguishable from a side-by-side high-resolution reference.

Relevant to our stuff is that they find that sharpening can (to some degree) compensate for real resolution.

Ah yes, this is pretty close to what I had in mind (I'd prefer a "blind" test, where the subjects have to choose left or right rather than state when the difference is no longer perceptible, but I have no idea which direction this would affect the results, if either).

What do you make of the sharpening compensating for real resolution? It's interesting in relation to the question of whether aliasing should always be viewed as a bad thing.

Quote
If your data contains a bandwidth >= fs/2, then you need to reduce that bandwidth in order to do a sampling according to Nyquist. In text-books this is carried out with a perfect sinc brickwall lowpass filter. If you dont do this, you will have aliasing.

OK, thanks for the explanation.

Quote
A valid point. But if we are exploring the limits of how close viewers may possibly get to the large print to investigate detail, may we not get into similar projected pixel angles?

Sure. But that opens up the (possibly interesting) question of whether what's best at a "reasonable" viewing distance will stay best at arbitrarily close distance.

Brian
Logged
opgr
Sr. Member
****
Offline Offline

Posts: 1125


WWW
« Reply #25 on: February 18, 2012, 11:27:12 AM »
ReplyReply

All I meant was that the typical font is viewed at a much lower resolution than the typical photo (and is obviously composed quite differently), so the aspects of human vision most relevant to the font designers could be quite different from those most relevant to photographers.

Brian

Actually the typical font is viewed in a much higher resolution than the typical photo, unless of course you haven't picked up a magazine lately, or watched printed material because the internet is the viewport of preference, and printed material is sooo last century…

I would like to suggest a possibly interesting simple practical experiment for more insight in this discussion:

Create a completely white image, and draw a single horizontal or vertical black "hair-line". (i.e. a single row or column of pixels). Then view this image at 100% magnification on screen. Move back from your screen as far as you can until you can no longer discern the line on screen. (Perhaps try different contrast settings). You could do the same with your printer. At it's highest native resolution, say 1440 dpi, print a single line at that resolution, and see how far away you need to be to avoid the nuisance of seeing a hair in your soup…

Our perception is very sensitive to edges. So much so, that a healthy 25 year old person with good eye-sight, when looking at a quality magazine at arm-length, should easily discern the difference between a font printed from a 1200dpi print-press-plate, or one at 2400dpi.

There is one aspect in all of the discussions about sharpness, resolution, and aliasing, that I miss. I haven't been able to quit find the right words to explain, so my apologies in advance if this seems somewhat Ray-esque:

There is a certain aspect in our perception of images that goes beyond the conscious perception. Those qualities of the reproduction namely that give an image a certain comfortable viewing experience. "Easy on the eyes" so to speak. Compare for example an overprocessed, over-sharpened image, versus a slightly blurry image. Even though the former may have all the measurable characteristics of more apparent sharpness or even real sharpness, the blurry image may still be more pleasing to look at.

Another simple experiment possibly worth a try:

Print a block of text at 300dpi with anti-aliasing.

The softness of the letters should be visible during reading. While the softness can be easily ignored, for some reason reading such a text is far more tiresome than reading sharp text. Which brings me to the example of text on screen 2007 Windows style or Mac style: there may be advantages to not using anti-aliasing to increase apparent micro-contrast useful for reading, but there is also a certain advantage in aligning all horizontals or verticals when appropriate, even at the expense of sharpness, because our eyes/brain will be more comfortable reading/scanning such lines. The question then becomes: How relevant are these trade-offs under certain conditions, specifically for imaging?

Architectural images versus Landscapes?





 
Logged

Regards,
Oscar Rysdyk
theimagingfactory
bdosserman
Full Member
***
Offline Offline

Posts: 120


« Reply #26 on: February 21, 2012, 09:34:35 PM »
ReplyReply

Actually the typical font is viewed in a much higher resolution than the typical photo, unless of course you haven't picked up a magazine lately, or watched printed material because the internet is the viewport of preference, and printed material is sooo last century…

OK, but you know perfectly well that the context of the discussion was screen font designers in comparison to photo printing techniques.

Quote
There is one aspect in all of the discussions about sharpness, resolution, and aliasing, that I miss. I haven't been able to quit find the right words to explain, so my apologies in advance if this seems somewhat Ray-esque:

There is a certain aspect in our perception of images that goes beyond the conscious perception. Those qualities of the reproduction namely that give an image a certain comfortable viewing experience. "Easy on the eyes" so to speak. Compare for example an overprocessed, over-sharpened image, versus a slightly blurry image. Even though the former may have all the measurable characteristics of more apparent sharpness or even real sharpness, the blurry image may still be more pleasing to look at.

This is absolutely right, but I think that the sharper image could also be more pleasant to look at, depending on the particulars of the image. This is the sort of thing I could imagine being at play when people claim to prefer a MF image in a small print as compared to a lower-resolution print. But again, it would be nice to have some blind comparisons to back this up.

Quote
The softness of the letters should be visible during reading. While the softness can be easily ignored, for some reason reading such a text is far more tiresome than reading sharp text. Which brings me to the example of text on screen 2007 Windows style or Mac style: there may be advantages to not using anti-aliasing to increase apparent micro-contrast useful for reading, but there is also a certain advantage in aligning all horizontals or verticals when appropriate, even at the expense of sharpness, because our eyes/brain will be more comfortable reading/scanning such lines. The question then becomes: How relevant are these trade-offs under certain conditions, specifically for imaging?

Architectural images versus Landscapes?

Yes, these are interesting questions.

Brian
Logged
hjulenissen
Sr. Member
****
Offline Offline

Posts: 1678


« Reply #27 on: February 22, 2012, 02:57:32 AM »
ReplyReply

This is absolutely right, but I think that the sharper image could also be more pleasant to look at, depending on the particulars of the image. This is the sort of thing I could imagine being at play when people claim to prefer a MF image in a small print as compared to a lower-resolution print. But again, it would be nice to have some blind comparisons to back this up.
The BBC whitepaper suggests that using a slightly sharper scaling function, you can "get away with" slightly lower resolution before visible errors appear compared to a flawless reference.

I am sure that size, distance, motif, viewer preference/background affect this to some degree.

If your image contains a fence in the horizon at a constant distance from the viewpoint, no-one will probably notice if it is rendered at 1 pixel cycles if it should have been rendered at 0.9 pixel cycles. In this case aliasing ("right value, wrong place") may be beneficial. If the fence is tilted compared to the camera axis so that it "sweeps" from 0.5 pixel cycles to 10 pixel cycles, then aliasing may produce visible patterns that the viewer finds highly unrealistic and annoying.

-h
« Last Edit: February 22, 2012, 03:04:06 AM by hjulenissen » Logged
Pages: « 1 [2]   Top of Page
Print
Jump to:  

Ad
Ad
Ad