It's Bigger, But Is It Better? Part II

Photoshop Bicubic vs. Genuine Fractals vs. BlowUp

By Ctein

My previous post made clear that none of these three upsampling methods is consistently superior to the others. Perhaps we can learn more of their behavior from other test cases.

But first...Lubor asked if incremental upsizing with Genuine Fractals produced better results. The illustrations above shows what happens when I upsized to 10X in five steps instead of one. The result emphasized Genuine Fractals' flaws as well as its benefits. Edges were even cleaner and sharper with incremental upsizing, but fine detail and highlights became even more polygonal and less natural-looking.

On to new cases. This detail shows a portion of another ISO 100, low-noise photograph from my Finepix S6000fd. It differs from the photo I tested last time in that it lacks the sudden and extreme contrast changes of that Christmas tree photograph. At both 1.67X and 4X, Genuine Fractals came in first, Bicubic Sharper came in second and BlowUp placed third. The differences, though were not profound. At 10X, Bicubic won out, not because of problems with artifacts from the two plug-ins, but because they both suppressed the fine detail in the stand for trees, mistaking it for noise. Their images were less sharp than Bicubic's.

Okay, let's take a closer look at this noise vs. detail business. Above is a full frame photograph from my camera. The lighting conditions were extremely poor, so I had to use ISO 1600. Even in this reduced-size image you can see how noisy the original is. (By the way, most of the noise can be successfully suppressed with Neat Image, but that's not the point of this test).

The three 1.67X enlargements are nearly indistinguishable. Genuine Fractals looked marginally worse; it mistook some of the noise extremes for real detail and enhanced them, so its rendition has slightly worse grain than the other two.

At 4X there's nothing to distinguish the three versions. Edges are no sharper with the plug-ins; noise looks similar in all the enlargements. Take it up to 10X? Oh, the horror, the horror! As you can see in the closeups, the two plug-ins, having confused noise for image, have dutifully enhanced it in an entirely unattractive way.

OK, there's a clear lesson here—if you're intentionally retaining noise for aesthetic reasons, don't trust your enlargements to "clever" algorithms. Stick with simple Bicubic. If you're not, be sure to do your noise reduction before enlarging.

Next time it'll be film scans and some synthetic test images that show some of the interesting peculiarities of all three approaches.

____________________

Ctein

Comments

[David Bennett]

Hi Ctein,
Where you said " ... fine detail and highlights became even more polygonal..."

is "polygonal" a special term of art with regard to digital imaging? If so, what does it mean?

David

[Sean]

Have you tried SizeFixer?
http://www.fixerlabs.com/
It uses camera profiles and an iterative process so it is slower. But I think produces good results.

[Ctein]

Dear David,

"Polygonal" means polygon-shaped; e.g., having some number of straight sides.

Incidentally, one of my favorite Google search features is to type in "define: ***" where "***" is the word you want definitions for.

Not as much fun as reading the OED, but handier.

pax / Ctein

[michael]

There's also dictionary.com, and they even have a thesaurus.

http://dictionary.reference.com/

Much more fun than the OED, especially the micro printed edition.

[michael]

One more thing: you mention Neat Image. I use Noise Ninja. Can you point me to tests of the various noise plug-ins?

I myself have tested Ninja vs. Noiseware, and I think they are equal.

Your thoughts?

[David Bennett]

Thanks Ctein,

Yes, I am familiar with the word 'polygonal' but I wondered whether it had some particular meaning in the context of digital imaging?

For example, it could be that the effect of upsizing is to "polygonise' details due to some aspect of the upsizing formula.

Or, for example, it could be that the polygonal shape of the iris in the lens is the root cause of this happening, and the effect is more clearly seen once the image is upsized.

These are just guesses out of the myriad possible reasons for polygonals showing when the image is upsized and I am not saying that either of these is the actual reason - I just give them to explain what the interest behind my original question.

I hope it is clear now.

regards
David

[Carl Leonardi]

Ctein, When your testing with Bicubic, which Photoshop are you using and if it's CS3 is it the regular Bicubic or Bicubic Smother. Which is best for enlargements.
Thanks, Carl

[Adam Isler]

And, amazingly, online is better for looking up things you wouldn't expect. I was reading a Raymond Chandler novel, "The Lady in the Lake," from 1941 last week and kept coming up against the word "chypre." My wife didn't know what it was, it wasn't in my Merriam Webster Collegiate dictionary, so I pulled out the microprint edition of the OED and a magnifying glass and it wasn't there either!

Stuck it into Google and instantly Wikipedia gave me a full page on it, with citations going back to 1917. No idea why the OED didn't have it but I'm a committed "online-ist" now - and don't get me started about what happened when we insisted on my 8-year old daughter working on a recent research project at the local NY Public Library instead of entirely off the web...

Adam

[david emerick]

A bit off topic, but if you are on a mac and using a cocoa application (Safari). hover over a word and press command+control+d for any dictionary definition.

On topic, I find careful incremental enlargement with careful sharpening works best. Craftsmanship always beats software shortcuts.

cheers

d

[michael]

From wikipedia:

Chypre is a name used to describe a family (or concept) of perfumes, usually based on a top note of citrus and woody base notes, usually from oak moss. The word Chypre is French for Cyprus.

Thought I should mention it.

[Damon Schreiber]

So were you saying that bicubic was better than the plug-ins when dealing with noisy photos or just not to bother with them because the results are the same? I ask because the detail photos of the crab look nearly identical to me, especially Genuine Fractals and Bicubic.

In any case, I haven't yet seen any compelling reason to ever use one of these plug-ins. Perhaps in part 3?

[erlik]

On a tangent, Ctein, have you tried PhotoAcute Studio? It won't allow you to upsize a photo enormously as it can only achieve "superresolution" of less than twice the original and even then the highlights are clipped though not polygonal. But it is wonderful for noise.

Since PhotoAcute takes several photos of the subject (handheld, too) and combines them, the random nature of the noise makes it possible for the program to do a very nice removal. Very natural looking and better than anything _I_ achieved with Noise Ninja.

The downside is that it really works only with still subjects and that you have to convert RAWs into DNGs if you want to keep a greater bit depth. And the conversion is irritatingly time-consuming. Plus you don't have anything like control over its processes.

[Johannes]

Thanks for the great review.

One thing I was wondering about is the difference between Bicubic Smoother and Bicubic Sharper.

From what I have found it would be natural to use Bicubic Smoother on upsizing and Bicubic Sharper on downsizing but I have also read in Scott Kelby's CS3 book that he recommends increasing the resolution and using Bicubic Sharper when upscaling.

It would be nice to have some comparison on the effects of these different algorithms.

Johannes

[Martin Doonan]

Like many others, I'm sure, I've been doing a little of my own testing given that I'm a GF user. I surprised myself when I started to see the same results as Ctein (making me wonder why I picked GF in the first instance).

What I've been working on is combination enlargement - i.e. using a mixture of bicubic and GF (I'm on GF 4).

What I've found - using bicubic followed by GF is consistently giving better results than any single tool (even with multiple steps). Best when applied as equal percentages (i.e. each enlargement is the square root of the final ratio - for 8:1, that's 2 lots of 2.82:1). Then I get the detail preservation of bicubic with the edge sharpness of Genuine Fractals. It's getting close to good enough for an automatic enlargement routine.

[Carl Weese]

Ctein,

Are you using bicubic sharper because of tests that disagree with Adobe? Bicubic smoother is recommended for enlargement, 'sharper' is intended for reductions. I haven't done careful comparisons but find upsize results with 'smoother' remarkably good. As an aside, I find even downsizing with 'sharper' results in harsh rendering so I usually downsize with plain vanilla bicubic.

[Cassio Dorneles]

Another nice scale program is GREYCstoration (http://www.greyc.ensicaen.fr/~dtschump/greycstoration/). AFAIK there's no Photoshop plug-in version, but it is free and does a good job.

[Charles Maclauchlan]

I appreciate the work you're doing here and wonder if you are familiar with Mike Chaney's magnificent printing program Qimage? I am assuming that upsizing / sharpening is being done with the intent to print the image and that's why I ask.

[Ctein]

Sean: The claim that it uses camera information makes it unusual enough to be worth testing. The authors don't have Mac Universal Binaries yet, but I've put in a request for review software when they do.

I think I should say that in general I'm not likely to test people's recommendations. Testing an upsampling program is time-consuming and costly. I have to have a good reason to think it's worth my while.

Michael: I don't know where there are comparison tests of noise plugins. Readers?

David: Thanks for the clarification on your question. The polygonalization is an artifact of the algorithms. They're inclined to emphasize edges.

Carl: I'm using Photoshop CS3. As to which is best for enlargements-- Bicubic, Bicubic Sharper or Bicubic Smoother -- my experience is that it depends on the photo I'm enlarging. Contrary to Adobe's recommendations, though, I rarely find Smoother to be the best choice. Most of the time, it's an even split between Sharper and plain vanilla.

As I said in Part 1, the Bicubic differences are usually subtle compared the differences between approaches.

pax / Ctein

[Ctein]

Dear Damon,

Sorry for not being clearer. What I found with the noisy image was that at all enlargement sizes the plug-ins were either indistinguishable from Bicubic or were worse.

I have one more film scan test to do, but so far I'm seeing a certain lack of compelling reasons myself. If these were $50 plug-ins, that'd be different, but several hundred bucks?

I mean, I'll use'em for the small incremental gains, but I got them for free!

pax / Ctein

[Howard Cornelsen]

I've been taught to use Bicubic Sharper to downsize and Bicubic Smoother to upsize. Is that idea either wrong or out of date?

Thanks!

[Raymond St Arnaud]

Looking down the wrong end of the barrel.

As the purpose of the cited programs is specifically to make images bigger, it would seem to me that the assessment of the results should be on how the images look at a reasonable social viewing distance. I suggest the following standard from the PPA

http://www.ppa.com/files/public/JCmanuaNov07.pdf

where the judges are sitting behind a table that places the judges 6 feet from the exhibited 16x20 inch photograph.
(Note: I am not a member of PPA or PPOC.)

I choose to pass on the definition of viewing distance as noted in the Photo-Lab Index, (page 9-11 in my edition) as I don't think it relates to real world viewing in a gallery or home/office.

Dan Margulis, in his classes, likes to use an example from a Velazquez painting, early 1600's, where he illustrates how artists use a technique to enhance edge contrast to give the illusion of greater sharpness or acuteness. (Darks are darker, lights are lighter along adjoining edges)

This often results in less detail in that area of the painting.

The focus on minute detail is misdirected and runs counter to the point of the programs.

The point of this testing should be not on the small details seen from 6 inches, but on the effect on a viewer at a reasonable distance.

In scrutinizing the enlarged images, I found that the enhanced edge contrast cited in GF to be an asset, and resulted in the illusion of more detail. I refer specifically to the area on the 3 horned beast between its eye and the crown of its shield. There are subtle areas of different density, and it looks to my eye that the edge contrast of GF yields a better final effect.

I choose to be with Velazquez and painting history on this test.

Regards
Ray

[Ctein]

Dear Martin,

Hmmm, based on one test print, that *does* look promising!

If there's a Part 4, I'll look into it further for that.

Thanks for the suggestion.

pax / Ctein

[Ctein]

Dear Ray,

Please reread the first part of this article and you'll see that I said that pixel-peeping wasn't part of the agenda. In fact, the "10X" image samples I'm providing you folks are actually downsized -- they're effectively at 25% viewing scale. Just so's folks can't pick at them.

Remember, also, that I'm making prints-- I get to view them at all sorts of distances. I'm moving no closer to the prints than I need to to see any difference between the upsampling approaches.

At the scale of 6 feet distance from a 16"x20" print (and, I presume, proportionately further from larger ones), there is no visible difference between them. Quoting from part 1:

"While larger prints are often viewed from some distance, if your prints aren't being given a critical examination there's really no need to consider either of these pricey plug-ins. Stick with fast and free Bicubic."

In the case of the Spider Crab, there is a VERY narrow range of distances over which the GF enlargement looks sharper and the grain doesn't look massively uglier. Much closer than that and the grain looks a lot worse. Much further and you can't perceive the difference in edge sharpness.

Finally, one last self-quote:

"I must remind readers that the point of illustrations is to illustrate my points, not to prove them. If you don't agree with my description, based on what you're seeing on-screen, trust the description."

pax / Ctein

[Robert Roaldi]

I have an idea for equipment reviews for Ctein. The infinite pixel detail deconstruction found on many other sites is just becoming too much for me. They are boring and I never read them in their entirety. Lots of data, little useful info, as I have complained about in the past.

Let me describe what I mean with an example. I bought a Sony R1 based on one line in the camera review on Luminous Landscape. At one point, MR tested the camera at 400ISO, said its noise was worse than whatever Canon D-SLR he was comparing it with, and then added that you wouldn't see the noise at prints sizes up to xx (where xx is some size that I can't remember right now but suited my purposes.) Good enough for me, I thought. One line told me what I needed to know.

So Ctein, how about this. You get companies to lend your cameras. You shoot some pics and make prints to the best of your ability (maybe even describe your procedures now and then but that's not necessary). Then you put a page up on the web saying something like, 5x7s are great, 8x10s are really good, 11x14s are ok, anything bigger looks like crap. I posit that almost all other info is superfluous. Of course, your judgements will be subjective. That's perfectly ok. So are Roger Ebert's about movies, but if I find that my tastes coincide with his, then his reviews will be valuable to me. Over time, people will calibrate your tastes with theirs. Your reviews will be short, sweet, to the point. You could always add a forum for the purpose of attracting web traffic so you can sell ads for lots of money. In time, you'll put all the other review sites out of business.

It's a slam dunk. :)

[Max]

Hey Ctein, since you seem to be very particular about texture in your images, I suggest you try something (if you haven't already). Convert any of these noisy images to LAB color in Photoshop and apply a gaussian blur to the color channels, a radius large enough only enough to break apart the color noise structure, leaving the luminance channel as it is (or sharpening it if you want to, but not to). I think the luminance noise alone looks great without the color blotches.
Or may I try it with one of your examples?