Has anyone measured printer rolloff?

[Doug Gray]

I've been exploring high DPI rolloff on my 9800. This is similar to MTF on camera lenses but is produced by the inkjet printer's algorithms in spreading different ink dots. It's somewhat variable. Black against white has pretty decent resolution but intermediate neutrals show significant high spatial frequency loss.

Specifically, my 9800, operated in the highest resolution mode with unidirectional printing and "fine" settings produces about a 50% reduction in signal at a spatial interval of 1/150 in when the image is 1440 PPI. This occurs when printing linear space sine waves with peak to peak "Y" range of .4 to .6 where "Y" is the reflectance. This is roughly a light gray. This is in the horizontal. I have not checked the vertical yet.

I noticed this when, after doing the large area reflectance fixup then scanning and comparing the scanned image to the generated proof image in Photoshop the sharpness of the scanned print was significantly reduced while everything else matched quite well doing A/B.

At first I thought it was the scanner's diffraction limits but it turns out the largest loss of resolution is in the printer itself. The scanner does have some spatial loss but it's quite small in comparison.

It's now clear why image printers that resample and sharpen are popular. I had not realized this was a real effect in the printer. At least with such a significant impact.

Any data on other, more modern, printers? How much does this vary with printer models?

EDIT-UPDATE:
I've since done more careful mesurements using cross correlation. The graph on this post overstates printer resolution. Particularly at high cycles per inch. because too much noise was included. Cross correlation has significantly more noise rejection.

[Doug Gray]

I've been looking more closely at printer "MTF" and ran across a paper that discusses various methods. Perhaps the most common one is the slanted edge technique which works well with purely linear systems such as camera lens, scanners, etc. Inkjet printers are odd animals because they must create a set of dots from a small ink set and spatially arrange them based on the RGB requested values at a fairly high PPI. This paper summarizes a few techniques and probably the closest to what I did was  Jang & Allebach's.  This has the advantage of looking at resolution using colors other than black and white. It goes of the rails a bit by discussing the non-linear effects (induced harmonics) from using L* and why one should use CIE Y. But hey, it provides background on CIELAB which may be of value to readers.

http://nicobonnier.free.fr/research/publications/EI2009_Lindner.pdf

I created 3, continuous 8 inch strips, .2" in thickness using sine waves of gradually increasing spatial frequencies from 3 cycles/inch to 300 cycles/inch. The 3 strips have a peak to peak Y range (Media white=1) of .33, .66, and 1.0. A segment was cross-correlated to the initial image, corrected for scanner attenuation (smaller than the printer MTF but still significant).

This was printed on Costco Glossy using the highest 9800 resolution with forward only printing. The image was created in MATLAB at 1440 PPI and saved in 16 bit space, linear gamma and printed with a custom profile made with just under 4k patches.



The three lines are yellow: Y=1.0 pk to pk, orange: .66 pk to pk, and blue: .33 pk to pk.

A couple points of interest. 20 cycles per inch is attenuated 10%, 100 cycles per inch is attenuated 40% and 210 cycles per inch is attenuated 90%

I also evaluated this printing the image at 90 degrees and expected a difference in resolution but the change was very slight.

So when you capture an image there are multiple sources causing loss of resolution. We are all familiar with lens MTFs and diffraction/motion blur. Here's something else to add to the mix.  They all add up and in the wrong direction.

These are the Windows 10 native driver results. I'm a bit curious as to how using Michael's QImage would modify these results.

Edited: description of blue and yellow were swapped. The yellow line is the resolution of the max swing (Y=0 to 1)

[NAwlins_Contrarian]

Okay you've caught my interest. But I'm too ignorant to have any confidence that I understand, and too tired to read that paper. So if you're willing to indulge me:

(1) Do I understand that this is cycles i.e. line pairs per inch? So 150 cycles on your chart means, e.g., 150 black lines interspersed with 150 white (unprinted) lines over the width of 1 inch?

(2) Does the Epson 9800 have a 'finest detail' mode, and if so, did you use it? I would have thought that at the regular 360 ppi, you would have hit a hard wall at that point--that even perfectly-placed lines would have maxed out at 180 cycles per inch--but again, I'm feeling quite out of my depth here.

(3) What software did you use to print, and maybe as important, did you just send the printer the 1440 ppi you mentioned, or did the software scale it to the printer's resolution, presumably either 360 or 720 ppi, right?

(4) I'll confess having a hard time visualizing your test pattern. Could you share at least a reduced-resolution version of it?

(5) As a follow-on test, and a companion to Jeff Schewe's photo tests in The Digital Print and also reproduced at https://www.digitalphotopro.com/technique/photography-workflow/the-right-resolution/, maybe you could do two versions, one with finest detail on and the other with it off, and compare the two.

Thanks!

[Doug Gray]

Okay you've caught my interest. But I'm too ignorant to have any confidence that I understand, and too tired to read that paper. So if you're willing to indulge me:

Well, since you have asked good questions ….

(1) Do I understand that this is cycles i.e. line pairs per inch? So 150 cycles on your chart means, e.g., 150 black lines interspersed with 150 white (unprinted) lines over the width of 1 inch?

150 cycles per inch refers to sine waves. See the image. This is the same as lines per inch but without the harmonics that black/white produces. Also, black/white bars alias pretty badly at the higher densities. The intent is to determine how rolloff occurs in the printed image.

(2) Does the Epson 9800 have a 'finest detail' mode, and if so, did you use it? I would have thought that at the regular 360 ppi, you would have hit a hard wall at that point--that even perfectly-placed lines would have maxed out at 180 cycles per inch--but again, I'm feeling quite out of my depth here.

Yes, it does and that's what I used.

(3) What software did you use to print, and maybe as important, did you just send the printer the 1440 ppi you mentioned, or did the software scale it to the printer's resolution, presumably either 360 or 720 ppi, right?

Photoshop CC.  The printed image is as 1440 DPI but I believe the printer may only be using 720 and tossing every other pixel. Can't really tell given how early the signal rolls off.

(4) I'll confess having a hard time visualizing your test pattern. Could you share at least a reduced-resolution version of it?

This image is reduced 2x from 1440 to 720 DPI. It shows significant aliasing in this post because it's downsized for display but if you load it in Photoshop and view it at 200% the aliasing is minimal. But I wouldn't trust this image beyond about 150 cycles per inch.

(5) As a follow-on test, and a companion to Jeff Schewe's photo tests in The Digital Print and also reproduced at https://www.digitalphotopro.com/technique/photography-workflow/the-right-resolution/, maybe you could do two versions, one with finest detail on and the other with it off, and compare the two.

Thanks!

As an aside, I started looking into this after the work I did on the scanner to mitigate/remove re-reflected light. My goal was to:
1. Print an image using Perceptual.
2. Make an image of what was presumably printed by converting to Perceptual in printer device space then converting back to working RGB using Abs. Col.
3. Scan the printed image, fix the reflected light problem with my program scannerreflfix.exe then load the scanned image in Photoshop (with high precision scanner profile attached).

Put both images in layers in Photoshop then compare the two. Properly done the images should match closely. And they did, except that the scanned image needed a significant amount of additional sharpening. I thought it might be the scanner and a small amount is. But it became clear that the printer itself was introducing most of this low pass effect. Hence the test and a quick scan of the rather limited literature on printer resolution.

[Doug Gray]

I ran the image w/o "fine detail" and indeed, the attenuation increases to 98% at 180 cycles per inch which corresponds to 360 PPI sampling rate. That's consistent with Jeff's article.

The 90% attenuation drops to about 150 and drops further to 130 when not using "fast" which prints in both directions.

Attenuation at 50 and 100 cycles is also very slightly greater, a few percentages.  Very consistent with Jeff.

[langier]

The bigger question beyond the quantification though, are you happy with the prints you get with your 9800? Charts and measurements only delineate the technical ability of the printer, ink, paper, settings. Sure, there's correlation with the quality of the final images, but that image is much more important, IMHO, than any measurements. If H.C-Bresson had stopped to measure the detail, MTF, resolution, grain, exposure, he'd miss the decisive moment, an instant much more relevant to his vision than all the testing, physics, chemistry, resolution, and measuring could ever produce.

As photographers, we like to get lost into these metrics. Yet Van Gogh, Van Dyke, Monet, Cezanne, Picaso, ad infinitum, were into the rendering of their mind's eye, damn the tools and measurements! With contemporary artists, I've never overheard a discussion of the canvas, brush, pigment nor pen in discussing their work or techniques.

It is interesting to find out the metrics of our tools and techs, thus you caught my curiosity and added to the sponge between my ears, but when all is said and done, it's the photograph that's surpasses it all! :-)

[John Nollendorfs]

When discussing photography, I find it interesting the discussions on pixel peeping, regarding new camera sensors, etc. , but rarely do the discussion migrate down to the level of the print in comparing the merits of new gear. A photograph is only a photograph when someone prints it! Do all the improvements of digital photography find their way into the print in a meaningful way?

I like to tell interested people, that when it  comes down to the  print on the wall, rarely can you ever see by looking, if the image was done with a $300 cell phone, or a $40.000 100 MP medium format camera. Bringing the digital image back down to an analogue print seems to mask all those differences we seem to see in pixel peeping comparisons.

[NAwlins_Contrarian]

Doug, thanks so much for the explanation, and for posting your test pattern. The use of sinusoidal oscillations between white and 'black', instead of using lines, now makes good sense to me; I understand the issue of avoiding harmonics and aliasing. I downloaded your test pattern, examined it on-screen, printed it, and examined it on paper. The values in the test image I guess reflect encoding for relative colorimetric of your printer's Dmax (in your 8-bit TIFF, 'black' was RGB 40,40,40), so now I get that one.

Being of the view that this is a worthwhile experiment, I took your pattern and added notations of what I believe to be the number of pixels per cycle at various points, for my own use and reference. My expanded / marked-up pattern is attached (albeit as a JPEG instead of a TIFF to get around the LuLa file size limitations). I tagged the file 720 ppi because that is an even multiple of my little Epson R280's 360 ppi that fits comfortably on a letter-size sheet. Then from Lightroom (6.14), I printed set for 8 inches long, which should give precisely a 2:1 downscale of the original pixels; obviously this implicate LR's downsampling, but that's usually what I actually use, and generally I find it to be quite good. The print is on Epson Ultra Premium Glossy Paper, using Epson's canned profile, with relative colorimetric rendering intent (because your 'black' is not totally black) and "Standard" sharpening. Theoretically then I should be able to discern down to at or just above (Nyquist) 4 pixels per cycle on the source file / markings, which downscaled to 2 pixels/cycle on the print. My aging eyes are an issue, and at the moment my distance-vision-enabling contact lenses are in. But looking at the print with an inexpensive 22x loupe:
* at original pattern 10 pixels per cycle / actual printed page 5 pixels per cycle, the maximum contrast oscillations are well-visualized, albeit as something closer to lines;
* down to about original pattern 5 pixels per cycle / actual printed page 2.5 pixels per cycle, the maximum contrast oscillations are reasonably well-visualized as lines;
* at about original pattern 4 pixels per cycle / actual printed 2 pixels per cycle, the maximum contrast lines start showing artifacts; and
* at about original pattern 3 pixels per cycle and below that, I see stuff, but it is not the proper pattern, and I guess reflects some combination of what I think may be artifacts at the high-resolution / right end of your file and LR's downsampling.
This seems to suggest to me that at least with maximum-contrast detail, effective printing performance is pretty close to what the printer's 360 ppi specification suggests. Note that to my eyes, your lowest-contrast oscillations start to become indistinct below bout pattern 8 pixels per cycle / actual printed 4 pixels per cycle.

Note first that I want to re-examine the test print when my contact lenses are out--I may see more / better detail.

Note second that it appears to me--to be clear, as you warned--that your posted image gets artifacty at the highest resolutions. I'm also posting screen-captures at 800% of the region at and below roughly 3 original pixels per cycle and the far right (high-resolution) end.

[Changing gears a bit]
As for the suggestion that

rarely do[es] the discussion migrate down to the level of the print in comparing the merits of new gear. A photograph is only a photograph when someone prints it! Do all the improvements of digital photography find their way into the print in a meaningful way?

I like to tell interested people, that when it  comes down to the  print on the wall, rarely can you ever see by looking, if the image was done with a $300 cell phone, or a $40.000 100 MP medium format camera. Bringing the digital image back down to an analogue print seems to mask all those differences we seem to see in pixel peeping comparisons.

First, I think that some people care and discuss that point, see, e.g.,
https://www.dpreview.com/forums/post/61644008
and
https://www.dpreview.com/forums/post/61634143.
Second, I do think the camera matters, but within certain limits of what print size and what camera, see, e.g., my post at
https://www.dpreview.com/forums/post/61545472.
Third, although I like prints, I also resist, in any but the most technical sense, the suggestion that it's not a 'real' photograph until there's a print--because I certainly have enjoyed, e.g., looking at transparency film on a lightbox, and in more recent years at digital images on a computer screen or large HDTV.

[Doug Gray]

The values in the test image I guess reflect encoding for relative colorimetric of your printer's Dmax (in your 8-bit TIFF, 'black' was RGB 40,40,40), so now I get that one.

Sounds like you converted it to some standard RGB space. The posted file is tagged with a gamma=1 RGB space and the "black" is indeed slightly less than black and runs about RGB=(6,6,6) in the gamma=1 space. And yes, it was adjusted above 0,0,0 to accommodate the printer's DMax. That's still about 98% absorbance though. I didn't want to drive the printer beyond it's limits and I was printing using Rel. Col w/o BPC.  Also, there is a small amount of noise (dither) added, about 0.3%, which I generally use for getting higher effective resolution using 8 bit images. Probably not all that significant, or useful, in this test.

Here's the frequency in cycles/inch from the left side of the 8" strip at every .5 inches.
0.0:    3.0
0.5:    4.0
1.0:    5.3
1.5:    7.1
2.0:    9.5
2.5:   12.7
3.0:   16.9
3.5:   22.5
4.0:   30.0
4.5:   40.0
5.0:   53.4
5.5:   71.2
6.0:   95.0
6.5:  126.6
7.0:  168.9
7.5:  225.2
8.0:  300.0

[Doug Gray]

Here's the full resolution (1440 PPI) file with tick marks and labels in cycles/inch.
Minor tick marks are at 3 4 5 6 7 8 9 10 15 20 25 30 40 50 60 70 80 90 100 150 200 250 300
Major ticks at 5 10 20 50 100 200



Note the image is in Adobe RGB with gamma modified to 1.0


Update: Link to tif file. This has been converted to ProPhoto RGB (gamma=1.8 ) which has better 8 bit characteristics than sRGB or Adobe RGB.
https://1drv.ms/u/s!AoQDySMc0uNih51XfgD9y_5gWyHUSA

[NAwlins_Contrarian]

Doug, thanks so much for the full version with the spacings marked. This will make it much easier to use. Also, even at 800% I don't see the high-resolution-end artifacts on this one that I thought I was seeing on the previous one.

When I go to save the new version from the LuLa page, the file says .tif but the browser thinks its a JPEG. I saved it as both, and both seem to work, but I'm not sure what to make of this. Also, I don't have Photoshop, so I've examined the files in GIMP 2.10.6, which I find easier for some things along those lines than Affinity Photo, but the test print was from Lightroom (6.14). LR seems pretty good about property decoding stuff in unusual encoding, and the prints looked right to me, but I can't promise you it dealt with the gamma correctly.

[Doug Gray]

Added a link to the tif file 2 posts back. Not sure why the embedded is showing as jpg. It also has it's DPI set incorrectly.

[Doug Gray]

This is a portion (about 1.8") from a 5x7 at 720 PPI print of the Kodak PDI Colordisc as modified by DryCreek. It contains the "proof" image of a Perceptual Intent together with a V850 profiled scan at the same resolution of 720 PPI. The print was made using highest resolution setting and fine detail which also prints at a sample rate of 720 PPI. The scan does represent what the printed image actually looks like with significantly rolled off (attenuated) high spatial freq. details.

Click on the gif file to expand it. It should show the proof, then the scan.

The row of bars in the lower left under "2" corresponds to 160 cycles/inch. It's not visible at all when printed without fine detail.

Note that 5x7 is a pretty small print. To compare on screen proof with a print you need to look at the prints form a short distance and the screen (at 100% zoom) from a much further distance. Doing that shows the scan is very close to the actual print. This shows the printer spatial rolloff. The difference between screen sharpness and print is much more apparent on these newer 4K screens.