Best Interpolation Process Qimage vs Photoshop - Poll

[Ron AKA]

The
interpolation method used for both the downrez and the up rez were
the same.

That is unlikely because the terminology for reduction methods is
ambiguous. For example, there are at least three ways to do
"bicubic" reduction. The most popular method, and probably the
best, uses a bicubic convolution kernel that is expanded in inverse
proportion to scale. Another method will first use some low pass
filter followed by an unexpanded bicubic kernel. Still another,
the worst, would be to just use an unexpanded bicubic kernel. The
point is that you can't depend on a name to unambiguously describe
the size reduction method, particularly in cases such as QImage's
pyramid which have no real reduction counterpart.

Also, even if the names were unambiguous, you do not want to see
the effects from different reduction methods confuse the
enlargement results. Thus, I would prefer to see just one reduction
used as a test image for all enlargement methods.

All I can say is what I did. In Photoshop I just used Image, Resize, Constrained Proportions, allowed Resampling and selected one of the Interpolation methods. They are the same for downrezing as uprezing. No idea what the software does behind the scene.

In Qimage I was using Print to File to get the small size print and maintained the 72 ppi. The same interpolation choices are there in Qimage for printing smaller as they are for printing larger. Again no idea what the software is really doing after I make my choices.

I did consider downrezing them all with the same interpolation process, as that would have been easier. I did some quick testing using nearest neighbour down and bicubic up compared to bicubic both ways and I thought I could see a difference in the final product. In other words the downrez interpolation had an influence on the final image. After that I decided to use the same process down as up to eliminate another variable.

Still looking for your choices?

Ron

 

[Ron AKA]

fotographer,

I certainly respect your opinion and what you see on a print. However, I still suggest the purpose of this comparison was to isolate out the effect of the mathmatical interpolation process. Doing this on screen allows you to actually see a 72 ppi image at full size, and I think quite easily see the large differences between the methods. When you print then this image is changed again by interpolation up to the print driver resolution using either the print driver, or manually in PS, or automatically in Qimage. This is going to introduce another major change to the image, that I was trying to avoid. I would suggest this is another essentially independent issue - the value of interpolation up to the print driver resolution and the quality in which this is done.

Ron

 

[Mike Chaney]

I would tend to agree, if all you are trying to evaluate are the interpolation methods themselves. Wait... did I say I agreed with Ruzinsky?

That said, this is an interesting exercise for reasons other than just evaluating the interpolation algorithms. By downsampling and then upsampling again, you might be able to get a feel for how good a certain program is at both resizing tasks: downsampling and upsampling. With high res cameras and the relatively small prints that people often print, downsampling is just as important as upsampling these days. If you use software A and software B to do a downsampling/upsampling combo and compare the results, you may be able to get an overall idea of how good the software is at sizing images regardless of whether you are going up or down. Some software might be good at upsampling but leave aliasing artifacts when downsampling. Other software might be good at downsampling without aliasing, but might not be so great at upsampling. Admittedly, your average user would need a bigger magnifying glass to see downsampling artifacts as compared to upsampling artifacts.

--
Mike
Author: Qimage, Profile Prism
http://www.ddisoftware.com

 

[Ron AKA]

Just as a note, the final JPG save for all the images was done in Photoshop using the Baseline Standard, and not the Optimized for the Web option.

Ron

 

[The Davinator]

The other three turn me off fairly equally because of artifacts.

Kevin

....On screen or in print? If you are looking on screen at the oringinal samples, you're doing nothing more than pixel peeping.

 

[DRG]

With the one hot thread of late, perhaps there is no appetite for
this, but as an engineer the subject facinates me. It is a
formidable challenge to take a limited sample of data and try to
recreate the original image. Purists were highly critical of
digital music when it first came out ... and perhaps some still
are, but one cannot argue that with correct sampling, Dolby and the
like that there are not some pretty spectacular results.

Recreating sound from a digital sampling of audio is actually not an appropriate analogy for upsampling an image. In the case of audio, the digital samples contain complete information for reconstructing the original analog audio signal (assuming the original signal was properly low-pass filtered to suppress frequencies above 22KHz to inaudible levels). And Dolby has nothing to do with sampling and refers to a company that licenses several different types of technologies related to noise suppression and surround sound. In the case of image upsampling, you're creating information without a well-defined goal beyond the final result subjectively looking good (which usually means maintaining smooth, sharp edges).

David

 

[The Davinator]

" In the case of audio, the digital samples contain complete information for reconstructing the original analog audio signal "

David,

You wouldn't say that if you compared a 24 bit 192kHz recording vs a CD at 16 bit & 44.1kHz sampling. The complete audio signal is NOT achievable within the CD Redbook Standard. As well, CD players that upsample bit rate and sampling frequency can create a larger sample to work with & thus assist larger bit rate DA convertors to work more accurately with the original 16 bit signal. But the audi vs image sampling is not a good analogy.

 

[fotografer]

Ron,

Perhaps I didn't make it clear. My printer native resolution is 1200dpi. But it's able to print at hardware resolution of 600dpi. Thus I took the pain to make sure I don't introduce any uprezing or downsampling by printing those images at 600ppi.

Anyway, I have given by on-screen preferences. So there...

--
fotografer

 

[Ron AKA]

Sounds like your printer has the same choices as mine. The problem is that the image is 72 ppi, and there is no way to print at that resolution. Even in draft mode my printer driver ups the resolution to 150 ppi.

Ron

 

[fotografer]

Ron,

That's why I said my images are teeny-tiny. Because in order to print at the printer resolution without any more uprezing or downsampling, the image has to be 'scaled' down to the equivalent size at 600 ppi, if you know what I mean...

--
fotografer

 

[Sabinhagen]

If you use the context of the comparison to be a computer science sort of evaluation, then many of the knocks against this comparison noted above would be invalid.

The only technical issue with the comparison is that process A may work best on this image, but might be near the bottom for images with much different qualities. But perhaps someone brought that up as I did not read all the responses.

The original poster, perhaps out of academic curiosity, suggested the poll. Most posters either missed the boat on what and why the poll is doing, or chose to essentially respond with something along the lines of "who cares - it is not a practical comparison". So, as predicted, nobody really was going to play the proposed game.

Likewise, saying the audio analogy is not a good one exemplifies missing the intent of the original post - I think it works quite well as an analogy myself. But now we are anal retentive about analogies!

--
Scott

 

[aruzinsky]

Still looking for your choices?

Ron

No ranking but I will give you two bests using different criteria.

Artistically, F is best because it is the only one that does not show jaggies and jaggies are aesthetically intolerable to me (Unless it is a picture of something like a saw.).

Nonaesthetically, as in forensic work or court evidence, A is best because it shows the most real detail. If you look carefully, it captures the most outer ripples.
--
Author of SAR Image Processor and anomic sociopath
http://www.general-cathexis.com

 

[Ron AKA]

Try sampling a 20 kHz signal at 15 kHz and then make it sound like 20kHz. That is what these interpolation methods are trying to do - make up for missing data points.

Your choice on the best method?

Ron

 

[sting]

Properly executed and sharpened images using Pyramid, Sync 256,
S-Spline, Bicubic smoother , etc, etc. etc, will have only the most
minor differences on print. I give the nod to properly executed
bicubic routines along with S-Spline & sync 256 over lanczos,
vector, and pyramid. But as I said, these differences at even

Which program offers sync256 ? I was headed to making an image interpolation project using sync interpolation using the programmable shader engines of the latest crop of nVidia GFC cards, but I would have guessed that the long filter response of a sync function interpolator would be unacceptible for pretty much any reasonably sharp image details. If an image has no sharp details, then downsampling/anti-aliasing filters aren't needed.

 

[Donald Cooper]

Image A will look best printed because it has had the living daylights sharpened out of it, and that ALWAYS works best for printing. The general rule for printing is to sharpen until the acutance needle goes well into the garish zone. Personally on screen I like C or F.

But what exactly does this have to do with anything? That's the bit I'm not understanding. Different algorithms work better with different types of images and on different types of output devices for sure. And high end print drivers (medical imaging usually) actually do that -- they analyze the image and then select the best uprez algorithm for the particular device at hand. But that won't "prove" anything about the relative merits of one program or one algorithm over another because they're all image and device dependent.

FWIW I don't use Qimage and I don't think it will give you a better print than any other program. But it has some very nice features that might serve some people very well. So if you can use those features you should definitely buy it. It's very inexpensive. And if you think the images look better printed then so much the better. Placebo is, after all, the most powerful medicine.

To me the more interesting test would be to debunk the idea that you need a specific number of pixels as an input to the driver/RIP. Print some with a ppi of 359 and some with a ppi of 360 and see if anyone can spot a difference. That would be entertaining.

 

[aruzinsky]

A captures at least one more outer ripple than F sharpened and therefore A is better suited for forensic work. Also, F sharpened seems to have more distortion at top center outer ripples.

There are several reasons that you might choose one method over
another in print versus screen. One major factor is final print
sharpening. If you take an image that has been sharpened for the
screen and dump that to the printer, it will not have the proper
sharpness because sharpening at 72, 96, or 120 PPI for a screen
will appear quite different when you print at 600 or 720 PPI. A
smoother image with less artifacts can handle a lot more
sharpening without falling apart, and the print process in general
can often handle more sharpness than the screen without visible
artifacts.

So a smooth interpolation algorithm coupled with a smart print
sharpening routine produces the best results. This is why Qimage
has smart sharpening for the final print that factors in things
like how far you stretched the original (how much it had to be
interpolated), the PPI being used by the print driver, and other
factors so that the sharpness of the print is consistent and
comparable to what you see on screen. The problem here, and the
reason you might prefer F over A on screen but A over F in print,
is because the samples in this thread have already been upsampled
without factoring in the sharpening needed for the printer.

With this in mind, final print sharpening can be used to make F
look better than A regardless of whether you are displaying on
screen or printing. See the samples below. I took F and applied
some more sharpening, and now, as you can see, F is just as sharp
or sharper than A and will appear so on both screen and in print
without the aliasing artifacts. So when looking at the samples
here, you have to factor in both smoothness and sharpness as to
which is better for what purpose. Also keep in mind that the
dithering that is inherent in all inkjet printing often hides some
artifacts like aliasing so that they are not as noticeable as they
are on screen. For this reason, differences between interpolation
methods can often be less noticeable in print than they are on
screen. Of course, a lot depends on the printer, paper, and other
factors as well.



--
Mike
Author: Qimage, Profile Prism
http://www.ddisoftware.com

--
Author of SAR Image Processor and anomic sociopath
http://www.general-cathexis.com

 

[aruzinsky]

Also, the left and right sides of the dark area at the bottom has true jagged edges that F sharpened has smoothed, again making A better for forensic work.
--
Author of SAR Image Processor and anomic sociopath
http://www.general-cathexis.com

 

[Mike Chaney]

I see that. Entire ripple bands aren't missing, however, only a tiny piece here and there and that might be caused more by the antialiasing method on the downsample rather than the upsample. Antialiasing, while it can sometimes remove a little detail that could be called "real" detail, also helps to prevent a lot of artifacts that create detail that shouldn't be there. Here is an example of a downsample from 1000x1000 to 400x400 by both PhotoShop CS2 bicubic sharper and Qimage pyramid with default settings. As you can see, the PS version created a lot of "detail" that is nothing more than aliasing artifacts. Of course, these are easily controlled/removed in PS by first doing a gaussian blur of say 1 pixel, but PS's default downsampling doesn't include the antialiasing. I think even for forensics, this is a tradeoff: is the extra tip of a ripple worth the invented "detail" that is just artifacts? Probably depends on what you are looking at/for, the image itself, and other factors. In forensics, you normally want all the detail you can get but on the other hand, you don't want to see things that aren't really there either.



P.S. in the spirit of the original thread, BTW, the only reason I'm not voting is that some may see my vote as biased.

--
Mike
Author: Qimage, Profile Prism
http://www.ddisoftware.com

 

[saynomore]

Placebo is, after all, the most powerful medicine.

This reminds me of an old Norm MacDonald joke. He would say at the end of his comedy routine: "Remember my motto: laugther is the best medicine... uh... unless you are a diabetic... then, insulin..."

It is a nice phrase, but it does not apply always, not even close. Qimage is probably not the best of them all, and certainly is not magic, but it is up there. Jaggies are pretty easy to see, they are not like a mild headache you think goes away because you put some titanium/manget patch on your forehead. What Qimage does most importantly, and people seem to miss this, is cut a lot of time, offering comparable quality to a longer and much more carefully done individual workflow.

As I said before, I recently had to print a list of more than a hundred pics that I'm selling prints of. I was not gonna print a whole set, so.. load qimage, drag the pics, set the size to 3"x2" and print in letter sized paper that's it. No need to write an action, resize or even arrange the photos in one letter sized paper. Resampling quality is not even an issue here, but that it saved so much time. I save files in png format instead of TIFF, and you can still print in any color space you want (png does not carry a color space tag), for every file because of the wonderful filters which include color space override/assign. Countless of bits of help that do not necessarily involve interpolation quality, which is quite good too.

To me the more interesting test would be to debunk the idea that
you need a specific number of pixels as an input to the driver/RIP.
Print some with a ppi of 359 and some with a ppi of 360 and see if
anyone can spot a difference. That would be entertaining.

This is another thing. Some very knowledgeable people, as you seem to be, don't seem to grasp this. It is not a myth. It is not a matter of sending the file at 359ppi and 360 ppi and not see a difference. Drivers work at some fixed PPI. You can test this in two easy ways. Find the PPI your printer works at. 600 PPI for Canon, and 720 PPI for Epson. A good hint is looking at the DPI your printer head works at. It will be a whole multiple of the PPI the driver needs, for reasons I won't elaborate but are somewhat obvious. Now...

First test:

Have a very low resolution image, say 100 PPI, with rounded high contrast and fine detail objects, and send it to the printer at that resolution. You will notice that there are probably jaggies. If the printer didn't interpolate, you would have blurry edges, but not jagged. The printer ink dots are too small to have created the jaggies, so the printer driver must be sending a jagged picture to the printer head. But your original picture you sent at 100 PPI was not jagged. You can conclude some interpolation is being done there.

Note that I'M NOT DISCUSSING QUALITY IN HIGH RESOLUTION PICTURES, as this seems to be the usual misunderstanding. Just the fact that interpolation is done at the driver. If you send higher res pictures you probably won't notice the difference.

Test Two, and easier: (DISABLE BORDERLESS)

Send a very large file to the printer like 600 MB or more, at any PPI other than the "native" PPI of the driver (600PPI Canon, 720ppi Epson). The spool will begin, and the printer will take a while to start printing after spooling.

After that, send the same large file, but change it to the native resolution. (you don't have to change the file in any way, just let it change the size in inches and it will balance.) The printer will start printing much quicker. This would mean that in the first case some processing was being done in the printer that isn't being done now. The only difference was PPI sent. Same file size, same pixel dimensions, (same file).

Sorry for the long post.

 

[saynomore]

not to take someone else's credit...