« What Lens Was It? | Main | Buy You A Present For My Birthday »

Tuesday, 26 February 2008

TrackBack

TrackBack URL for this entry:
http://www.typepad.com/services/trackback/6a00df351e888f883400e5508269258833

Listed below are links to weblogs that reference It's Bigger, But Is It Better?:

Comments

Ctein,

I'm curious about the scaling factors mentioned. Are you talking about 10X, 4X and 1.67X linearly or about an area? If it's a linear measure I'm curious as to the choice of 1.67X, it's neither 2x (1.414) nor 3x (1.732). Although to be perfectly honest, if it's an area 1.67X is still curious...

This is just intellectual curiosity on my part, but at the same time it might give a better sense of scale.

Ctein,

what magnification factors are you talking about? Are you referring to linear enlargement factors or to the number of pixels? And which bicubic resampling method exactly are you referring to? Plain Bicubic, Bicubic Sharper, or Bicubic Smoother?

Some people seem to believe Bicubic Sharper was the same as Bicubic, followed by a small amount of sharpening ... and accordingly, Bicubic Smoother the same as Bicubic, followed by some blurring. Not so! It's three different resampling algorithms. When comparing methods for upsizing, you should include all three members of the Bicubic family.

For linear enlargement factors between 1.4× and 3×, I am getting good results when first applying some gentle Sharpening for Source (as opposed to Capture Sharpening), then upsize using Bicubic Smoother.

-- Olaf

Hi Ctein,

I know how difficult it is to objectively test and compare products and techniques, so please take this simply as my thoughts as I read this.

1 - Why not use Bicubic Smoother, which is what Adobe (and I believe Thomas Knoll) would recommend for enlargement in PS, rather than Bicubic.

2 - By feeding the Epson at 300 PPI rather than 360, we are involving the Epson driver to interpolate the data into the native 360/720 internal resolution of the input stage of the driver. So rather than print the files directly, the driver must interpolate again from 300 (file) to 360 (Epson native). So best we can say about the test comparisons is "this is how the programs/techniques compared (judged by the hard copy results...the prints), when they are further interpolated by the Epson Driver. We have no idea if the results would be skewed if we fed the Epson its preferred native resolution" (obviously the files published in the post are not subject to this).

3 - Since in the end we are not interested in what the files look like on screen (except for intellectual peeping), it would seem appropriate to me to sharpen the files before printing as would be done in the real world. It could be that after appropriate sharpening the results would be skewed.

I hope this adds to the discussion. It is not meant as a criticism or diversion, just some points that flew into my mind when reading.

Thanks.

Michael Tapes

In my experience, different sampling programs work differently at the same scale, producing cartoonish results at different levels, depending on the original jagged lines' scales. Starting with a smaller pixel count file for a given image Genuine Fractals produces those very weird piramids and poligons while enlarging, but the same structures at smaller scales (larger initial images, in the detail/pixel ratio) produce super clean results that look great.

Ctein,

I have to admit that I was always fairly suspicious of the claims people were making about Genuine Fractals. But out of curiosity, do you have any sense of how your results would compare to printing the original image using Qimage?

Best regards,
Adam

A couple of years ago I read everything I could find on this subject, and did a lot of my own testing. I came to the similar conclusion that results from the different programs varied from photo to photo with no clear cut winner. Since I never print larger than 16x20, I concluded that none of the 3rd party programs were sufficiently better than Photoshop bicubic to justify their cost, so that is what I've stuck with. I'll be interested to see what your final verdict is on larger size prints.

Ctein,

Interesting! Knowing that you are also a PWP user, would it be possible to ask for a very small postscript comparing the various resize algorithms available in PWP? I have tried some of them but never really found one ot be consistently best.

Thanks
Andrew

I've been looking at these plugins in anticipation of making large prints from the Sigma DP1. The unique challenge there seems to be aliasing artifacts. I look forward to parts two and three! Regards, Amin

Another request for comparison to Qimage which is said to have its own uprezzing scheme.

Has the old maxim about upsizing using bicubic repeatedly at 10% increments finally been put to rest, then? I learned about it when I was first teaching myself Photoshop in the mid-90s. I guess I never thought to re-visit the question. It was genuinely better than standard one-shot bicubic in my tests back then, though.

Did you go from original to re-sized in one "hit"?

I have read that if going for extensive enlargement, it is better to do so in several steps??

Michael Tapes- Although the literature states the native resolution of the printer may be 360, Mac Holbrook (nash editions), Jeff Schewe, Vincent Versace and Bruce Frasier all swore by 240PPI when at the Adobe/Epson seminars. I don't know why they were so insistent, but they really seemed to like the results they got across a variety of different photography subjects.

Andrew, you mentioned PWP (Picture Window Pro). In creator Jonathan Sach's white paper "Image Resampling," he makes it clear than Lanczos 8x8 is slightly superior to Bicubic, whether the latter is sharpened or not, and my experience is in agreement. He also shows that Lanczos 8x8 is superior to Genuine Fractals (mind you, this white paper was authored in 2001). I in no way consider myself an expert, but I am very happy with the results I get with Lanczos 8x8.

I came to the same conclusion.

I bought a program from Calumet, from the Netherlands called S-Spline. I found that the bicubic in Photoshop was equal.

Live and learn. Sigh.

POWERSHARES DYNAMIC MID CAP VALUE PORTFOLIO
Packaging with Perfection
Parents Without Partners
PWP Wrestling
Peter Walker and Partners Landscape Architects
Pers' Wastaiset Produktiot
Professional Women Photographers
Pretoria Wireless Project
Partnering with Parents
PWP is an undergr0und [sic] hacker group
Public Works Productions

Andrew,

The above are the Google search results for "PWP". ;-) I'm going to go out on a limb and assume you weren't referring to any of the above. Can you tell me what PWP stands for?

Many thanks,
Adam

Ctein,

Excellent idea this test. Have you considered testing Qimage (as mentioned by Adam), Photozoom Pro and the Fred Miranda plugin as well, tested at the Epson native resolution?

Looking forward to the next post that will help us decide on the best workflow.

It would be interesting to see a comparison with prints done accoring to the latst theory from Schewe - that is don't upres but just keep the ddi between 480 and 180, sharpen properly fro that res and let the printer driver do its stuff. If res falls below 180 then upres 2x , sharpen properly and again print. ( see the camera to print vidoe on luminous landscape)

Keith

Dear Ctein,

Thank you. I look forward to the rest of your report.

For what it's worth, my own recent experience: Not long ago I was asked to make a 2x enlargement of a 10MP file in a hurry. I read lots of reviews and tests, and lots of "it depends" conclusions.

Came across Fred Miranda's method of using Bicubic in graduated multiple passes, which he says is how it was designed to be used, and which he felt preserved fine details better than Genuine Fractals.

http://www.fredmiranda.com/SI/

In a follow-up down at the end of that article, he finds that Lanczos interpolation is also quite good, and is available in the free program IrfanView, as well as in QImage.

http://www.irfanview.com/

Lanczos did indeed work best for my task, at least of the options that were affordable and readily available to me. The next headache was figuring out where in the workflow interpolation should go, especially w/ regard to noise reduction. (I'm sure "it depends".)

*I should note that since he wrote the article I reference above, Fred Miranda has further refined and automated his "Stair Interpolation" action, which he offers for sale on his web site.

Cheers!

There are two other software options I would suggest:

Photozoom Pro (http://www.benvista.com/photozoompro/) with its S-Spline interpolation.

I've used this for printing 5 megapixel images from my E-1 at 20x24 inches, and it avoids any overtly digital look (aliasing jaggies or the plastic look) in exchange for a slight watercolour effect.

It has also proven popular in the video realm for up-rezzing standard definition video footage to HD.

I have not used, but always wanted to try Qimage (http://www.ddisoftware.com/qimage/) and its Pyramid interpolation.

Would it be worth your time to include Qimage? It is much cheaper and has many different resizing algorithms to choose from.

I second the Qimage suggestion. I like it because it means I only have to produce one final sharpened file, and Quimage will scale and tweak it to the right final size.

It seems up-sizing is even more tricky and mysterious than I thought - and hoped.

Having neither a giant printer nor access to the more expensive plug-ins, my experience is slightly different.

For a friend, I tried to make modest size, 3x4"ish, if I recall correctly, print versions of some tiny web images.

Magnifications varied, but were up to maybe 10x. The bicubics and FM's SI Pro just made big blobs, useless. The trial version of Qimage simply amazed me, seemingly working like the magic on TV cop shows, pulling detail out of thin air and saving the day.

Needless to say, I bought Qimage.

At more modest magnification, 2x, I did pixel peeping comparisons of several options with the bicubics, FM's SI Pro and his camera specific Resize Pro.

Resize Pro was the clear winner, with less apparent noise and clearer detail.
http://moosemystic.net/Gallery/tech/Process/UpSample/Uprez.htm

Moose

Dear Folks,

Some quick comments (gotta lunch date):

1) All dimensions are linear. The only time people should use areal dimensions in photography is when they're discussing medium costs. Otherwise, it doesn't correlate with any photo characteristics.

2) Why 1.67X-- cause I scaled up my 'native' 300ppi images to 500ppi. Arbitrary choice-- close to going from 8x10 to 14x17. Midway between going up one standard size (e.g., 8x10 to 11x14) and doubling the size (e.g., 8x10 to 16x20).

3) Yup, bicubic, bicubic smooth and bicubic sharp do different things. Make less of a difference in output quality than one would think, though, for real world images (I've got some synthetic test target images where the differences are spectacular.)

The results were more subtle than the differences between approaches. But, yes, I did look at all three, and I just lumped them together, same way I lumped together different settings within the plugins. Picked the versions I liked best from the three approaches and compared those.

4) 300 vs 360 ppi output? No, it's not a crock-- I've got synthetic test images that clearly show a low level of artifacting from printing on the Epson at 300 ppi that aren't there at 360 ppi. But they're on the scale of single pixel widths, and they don't represent real photos.

I can't see any difference in real photos printed at 300 ppi vs 360 ppi, even in stuff that's tack sharp down to the pixel level.

For historical reasons, my stuff all comes out native at 300 ppi. Might be interesting to try upsampling it to 360 ppi, using the three methods, and seeing if the prints looked visibly different (better? worse? sideways?)

Be that as it may, any differences due to printing at 300 vs 360 ppi are MUCH smaller than the differences between (and artifacts produced by) the different upsizing approaches and do not affect the results I'm reporting.

pax / Ctein

Dear Folks,

Ok, lunch appointment hasn't shown up yet-- I'll squeeze out some more comments:

1) Sharpening for output has no bearing on these results. It's just another source variable. There are so many!

2) Interesting question about incremental bicubic upsizing. Dunno the answer. I may test it on the photo shown in this column for a follow-up "part 4."

Note: I ran my upsampling tests on more than a half dozen different kinds of images-- ain't gonna repeat that for any follow ups! I have neither the time nor money nor interest.

*IF* I decide to do a follow-up, it'll be on just this one photo, so mileages may definitely differ.

3) Haven't tried Qimage. Enough of you are swearing by it that maybe I'll hit them up for a review copy and run a follow-up test (ala comment 2 above). Or not [weary grin]. We'll see. No promises

4) Good question about Picture Window Pro. Problem is I've been running these tests on a Mac. Dunno if things will print out the same way under Windows. I *may* look into this, vis comments 2 and 3. Also no promises.

pax / Ctein

My preference is PhotoZoom Pro 2. I only do modest enlargements but it's the only program I have found that gives me results that are sometimes better then PS.

I always go for an even simpler solution, Photoshop fit to paper checkbox in the printer dialog. Works very good, and very fast.

Is there a free implementation of bicubic?* Besides the one that comes with Photoshop, I mean. Sure, it's "free," but only in the sense that you've presumably already paid for Photoshop.

Otherwise, I'm looking forward to parts 2, 3 and 4.

On another note, Ctein, have you or Mike given any thought to comparing stitching/panorama programs? Although I bought Autopano Pro last year, I still think this would be interesting.

* Ur, the bicubic family of algorithms...

I have conducted similar tests with the same products and came to similar conclusions. There is very little discernible difference between the three. I use Bicubic (sharper) for the majority of my work for clients as it gives great results with quickly and GF for exhibition prints enlarging D2 files to 1 meter. If I hadn't bought into the marketing of GF I would probably be happy using Photoshop for all my enlarging. As it is I spent a lot of cash on the plugin and I'm gonna use it!

For those that asked about the upressing in several smaller steps, I believe Jeff Schewe in "From Camera to Print" (or maybe it was somewhere else) said that in current versions of Photoshop, that is unnecessary. Apparently, it used to provide better results, but has been improved to the point where one step is just as good.

Moose,

It looks to me as though Resize Pro adds some sharpening and contrast enhancement. Look at the rocks in the water in the upper left corner. They seem to be getting brighter highlights than exist in the original. Also, the original is pretty noisy itself, so the reduced appearance of noise in the upressed image makes me wonder what else is going on.

This isn't necessarily bad, but it shows that there is more than just upressing going on. Of course, you could sharpen or adjust contrast after upressing with the other methods, too.

Best,
Adam

"Is there a free implementation of bicubic? Besides the one that comes with Photoshop, I mean. Sure, it's 'free,' but only in the sense that you've presumably already paid for Photoshop."

Anyone know if the latest Photoshop Elements (Win and/or Mac) have the same Bicubic upsizing capabilities as full Photoshop? That would qualify as close to free in my book, since Elements costs so little.

Mike J.

"Is there a free implementation of bicubic? Besides the one that comes with Photoshop, I mean. Sure, it's 'free,' but only in the sense that you've presumably already paid for Photoshop."

For Windows users - download Irfanview (including the plugin package) and PhotoFiltre. between them these two programs can give you almost all the photo editing power you'll ever need without spending a penny!

Cheers, Robin

The behavior of the Genuine Fractals makes sense if you keep in mind that it was originally designed as a file compression program that converted the raster information to vector information and compressed it using wavelets. It tends to want to define things as areas with edges. On the one hand it does a good job of getting rid of aliasing effects on edges and sharpening them without the usual sharpening artifacts, but at the expense of wanting to make things look like areas with edges if the blur is near the sampling frequency.

Moose: Cool comparison. Thanks! And I sympathize about the "set-it-and-forget-it... NOT" aspect of all this. I, too, was hoping for a simple answer and a quick, rote routine I could run before printing.

MWG: Not running Elements yet, so I can't say if it's got the same bicubic algorithms as proper Photoshop-- something you could ask in the Adobe forums.

But... regardless, both BlowUp and Genuine Fractals are Photoshop plug-ins that require either PS or PSE. So, yeah, baseline to this review, Bicubic *is* free. And if you don't have either PS or PSE, this review is worthless to you.

Craig: The nonintuitive result of this first test is that the more I was enlarging, the LESS useful GF was. Note that it only beat out Bicubic at 1.67X. By 4X, the artifacting was too annoying.

BTW, everyone, both BU and GF have free 30-day fully-functional trial versions.

An aside-- despite what Adobe recommends, I hadn't found in my daily work that Bicubic Smoother was consistently best for upsizing and Sharper best for downsizing. Often as not, I preferred one of the other options.

pax / Ctein

Robin: There's also GIMP. But the real question that was being asked was if PSE implements Bicubic the same as PS, and the same could be asked of Irfanview or GIMP. Remember that BC's not a fixed algorithm, it's a kind of curve fitting and the designer can weight the fit parameters differently.

MWG: I really like doing panos. But it's not central to my art/business, and I'm happy enough with PS CS3's photomerge function that I'm not likely to ever try anyone else' software unless someone's paying me to.

Photomerge was just plain crappy and useless in previous versions of PS. In CS3 it really works, even for a demanding pro like moi!

pax / Ctein

1.67 is a good one to try, but so is 1.23, 2.56 or anything else that is not a whole number. If you enlarge something 4 times, you just get the same pixels four times and all the software has to do is smooth out the tones and a bit of sharpening.

If you do anything else, it can still only add whole pixels at a time. So it might add a pixel for every 2 pixels. Now what color is that pixel going to be? And you are going to have to do anti-aliasing.

This is hard enough that photoshop doesn't even do it when zooming; if you zoom out your large file to 50% or 25% in Photoshop, it looks quite good. Do it to 16 2/3 % ("16.67") and it looks horrible.

In the real world, you are unlikely to enlarge by whole numbers only. More likely, you will enlarge whatever you have to fit, say, 16 x 12" in 300PPI.

So, for the rest of your tests, skip the 4 or 10 times. Instead, go for 4.2 and 9.7 - those are the hard ones where a smarter algorithm is going to make the difference.

Ctein, Just did a quick up rez using bridge and out putting a 10 meg 40D at 25meg output. did a 40X60 on screen of both 10 and 25 meg using bicubic smoother and the 25 meg looked a little better then the 10 meg.at 100%. I doubt at viewing distant you could see the difference. I think at viewing distance there won't be much difference with any of the programs. This is where 20 megapixels+ and the best lenses start to show there stuff.

Gimp has a pretty good Bicubic implementation and a variation on the Lancosz interpolator as well (the Sinc algorithm). In my experience with both, they are all but identical at small scale differences (like the x1.6 above). For large upscaling, bicubic will stay a bit sharper on edges, but with a bit of the jaggies and a bit more artifacts than Lancosz. Which is better depends on the image more than anything else.

For downscaling - to use an image on the web, for instance - the situation is different. Lancosz clearly preserves more detail and edge sharpness. It's really no contest there. If you downscale, use Lancosz.

Oh, and if you print, the idea of just not rescaling at all is not a completely dumb idea. The state of the art in rescaling is pretty well known, after all, and as Cteins excellent test shows, the difference between methods is pretty marginal. A quality photo printer driver probably has a pretty good set of algorithms built in already, and it has the great benefit of being adapted for that exact printer. I would not be surprised if that has much more impact on the final quality than anything you can do by rescaling yourself.

Dear Bas,

Photoshop does screen displays very differently than it does resizing. Screens are quick and dirty calculations, designed for speed.

Upsampling's entirely different. What you're describing is how 'nearest neighbor' works; it's how I made the first figure in the article. Bicubic, BlowUp, and Genuine Fractals don't work anything like that, and they do not favor integral magnifications.

Try it for yourself-- you won't see any difference,structurally, between files upsized by 4X and 4.1X (for example) with any of these. I guarantee it.

pax / Ctein

Did you also try using Lightroom for the interpolation. i believe it uses an implementation of the Lancosz interpolator (similar to what is used in Qimage) - and supposedly gives better results than bicubic. I could try it out but as I have a 3 year old running around and creating havoc in my computer room I don't really have thetime at the moment.

Steve

I don't think any comparison of interpolation methods can be complete without testing all the methods that SAR Image Processor http://www.general-cathexis.com offers.

Dear Eduard,

"... Photoshop fit to paper checkbox in the printer dialog..."

When I try it, the output is only nearest-neighbor resampled, which looks like crap. I can't find any setting controls that let me specify how the upsizing is done.What are you doing that's different?

pax / Ctein

Steve: I don't own Lightroom.

Andy: Who said anything about completeness [smile]?!

(BTW, their web site is down, so that URL does not work. As an aside-- I've noticed that a fair number of the URL's that people give here as referrals don't work. Folks, PLEASE test URL's immediately before posting them, OK? It'll save everyone else's time. Thanks!)

pax / Ctein

"The behavior of the Genuine Fractals makes sense if you keep in mind that it was originally designed as a file compression program that converted the raster information to vector information and compressed it using wavelets. It tends to want to define things as areas with edges. On the one hand it does a good job of getting rid of aliasing effects on edges and sharpening them without the usual sharpening artifacts, but at the expense of wanting to make things look like areas with edges if the blur is near the sampling frequency."

Hugh, thanks, that was very informative. It also matches my preliminary findings in using this plugin to upsize the DP1 sample images from Sigma. These images are prone to aliasing effects when upscaled, and GF seems to be very good at keeping them in check. I posted some examples here: http://aminphoto.blogspot.com/2008/02/upsizing-images.html

Regards,
Amin

How about Qimage? I am using this program now for over a year. It is cheap and does a wonderfull job of resizing your print. Without you having to worry about it. I have got the idea from the Epson wide format site, al lot of profesional printers use it.

rgrds

mwg,

Some additional "free" Windows OS implementations of the "bicubic family of algorithms": Internet Explorer 7 (display only of course), paint.NET, Windows Live Photo Gallery and *possibly* XP's Windows Photo and Fax Viewer (can not verify).

TWIMC,

My ancient Photoshop Elements 4.0 for Windows offers bicubic, bicubic smooth and bicubic sharp resizing strategies. AFAIK later versions keep these options.

Adam,

PWP - Picture Window Pro from Digital Light and Colour http://www.dl-c.com

A truly excellent editor for a fraction of PS price. Also, its author has an interesting past (Ever heard of Lotus 1-2-3?).

Ctein, I checked the URL :-)

Andrew

Ctein,

You're right that Andy's link doesn't work, but through Google's cache search, I was able to find this link: http://www.general-cathexis.com/interpolation.html which DOES work and includes mouse-over comparisons of 21 different upsampling methods (including several commercial methods).

The take-aways (for my purposes at least) are relatively simple:

(1) A few of the methods are obviously unusable and can be dismissed.

(2) Of the rest, no one method is a clear standout. Depending on which part of the sample image I look at, some are better than others, but they all seem fairly close.

The bottom line is that you just can't create information that isn't there in the first place. So we are left to trying to reduce the appearance of artifacts. Since the best way to do this will vary from image to image and even within various parts of the same image, it seems to me that you can either (A) drive yourself completely crazy by spending hundreds of dollars on plugins and spending hours upon hours comparing the results using all of the free and commercial methods available each time you want to print an image, or (B) just use bicubic (or whatever equivalent method is included in your favorite image editing program or print driver) and be done with it.

Unless parts two and three of your analysis include a revelation, I'm going to stick with (B). Just call me Dr. Strangelove.

Best regards,
Adam

Interesting that Ctein finds so many URLs that don't work, since I check them all (or almost all--I suppose I get lazy sometimes) before I post the comments that contain them. Then again, I'm not checking the links in the actual posted comments, I'm cutting and pasting the URLs from the comment emails as they come in to my email program. I wonder where the discrepancy lies....

Mike J.

I think you are defining a test methodology and issues to watch for, not answering anyone's specific, personal, individual questions. I just did a quick check with a friend who is preparing poster size prints from his K10D. The optimized methodology that works on his Pentax landscapes doesn't work nearly as well on some from my E-3 (left the M8 in the closet this time). Even though when the same originals are viewed at 100-200% the E-3 work was better in several ways. Don't know if what worked for the K10D will have to be reworked for the K20D.

scott

If the Genuine Fractals is better than Bicubic at 1,67X, and not at 10X, what about to use it in a few little steps to achieve the higher size? Maybe the result will still remain superior?

Ugh. Why is it that the Internet is so fleeting, and yet so permanent? It is so easy to dash off a comment, but so hard to correct it, edit it or pull it back!

The last part of my previous post obviously should have read, "Unless parts two and three of your analysis include a revelation, I'm going to stick with (B)." [Adam--I corrected it. MJ.]

Rereading what I wrote also led me to believe that the tone of my written comment is far removed from the tone intended when I wrote it. Text is a very poor substitute for the nuances of the non-verbal communication that occurs when we talk to each other in "real life". So, just to clarify...my comment wasn't intended to be an attack, or a critique, nor was it intended to belittle Ctein's analysis and efforts. In fact, I was trying to say something along the lines of "Right on! I agree with you wholeheartedly and here is some more information that may back you up." [I only capitalized "DOES" so that people would see it and not assume that the link was broken.]

At any rate, I look forward to parts two and three of this article, and I am very grateful to Ctein for "spending hours upon hours comparing the results" so that I don't have to!

Best regards,
Adam

P.S. Andrew: thanks for the explanation. Do you have a sense of the relative merits of Picture Window Pro vs. Paint Shop Pro Photo X2?

Digital Outback had an uprezing contest a few years ago

http://www.outbackphoto.com/contest/contest_01/essay.html

and, I think, some additional ones later.

The link includes contest entries and workflows -- some of which appear almost painful!

From the comments, I gather that Ctein and others, like myself, have found that getting a bead on the best way to upscale digital images is not simple.

Dismaying, especially since
while shopping cameras, I'd been leaning toward the idea that upscaling from fewer and better pixels would be at least as good as having more pixels to start with. Drudgery adds a caveat to that theory.

On the other hand, I find perversely comforting the fact that, despite all the computational magic available to us, getting the best results depends on applying human experience, intuition and judgement to each case.

The different opinions, observations and test results for various upsizing algorithms is to be expected.

Upsizing attempts to get something from nothing. When an image is upsized, data (pixels) are added where no information exists. This means the added data are constructed from parameter estimates for the real, but unknown, data values had they been collected. The uncollected data values are real in the sense that if a sensor with twice the pixels had been used, there would be empirical values for the desired additional pixels instead of estimated values. Of course more empirical pixels don't always mean better pixels, but upsizing attempts to model a sensor that has the same pixel quality as the actual sensor, with many more virtual pixels.

While the parameter estimates used to generate the virtual pixels are computed using mathematically valid principles, they are essentially guesses. Like all guess, the value of the guess depends on prior information and the model for the data.

Parameter estimates require a model for the data. It is reasonable to assume that different types of photographic images can require different models. More than one model can account for the data, but which one best represents the data?

DIfferent people can upsize the same image with the same application (algorithm) and get different results if they use different input parameters (prior information). Using completely different algorithms (application) only leads to more diversity.

What we know is:

o A moderate, well thought out degree of upsizing can improve the visual impact of an image

o There is not a great deal of difference between approaches for most images

o When there is a significant difference, it is due to either how the photographer(s) selected the algorithm input parameters, or how well the application models the data in a particular image

o There are no miracles. The data is.

The benefits form upsizing are always limited. A large increase in image size from a single calculation is highly susceptible to the assumptions used in the calculation (model for the data and algorithm input parameter selection). Some algorithms benefit from a a series of small upsizing calculations because the cumulative errors are less than the errors from a single calculation where many pixels are added at once. But after the first upsizing calculation, subsequent calculations are adversely affected by errors in the previous result.

o The more information (i.e. closely spaced lines) present in the original image, the more important upsizing becomes... as does the data model and input parameters. Images (or regions of an image) that do not have high information content (sky, out-of-focus objects) require little improvement and upsizing results are less sensitive to variables.

It is difficult to create something from nothing.

Ctein:

Good work.

What I'd be interested in NEXT is a review ( or even a few choice comments) of ViaVoice for mac which you mentioned somewhere and a comparison with iTalk for mac.

Regards

Uhhh... that last post should refer to iListen, not iTalk, for mac. The only two voice recognition programs for mac as far as I know.

I'd especially be interested in how these programs work with digital photo editing and viewing programs.

Thanks,

Jeff Glass

Amin: Thanks for posting those comparisons. I don't have access to Sigma, and the fine image structure from those cameras is different from standard ones. By the way, for the benefit of readers, the magnification you used is very close to my 1.67X, so your tests should be lumped in with those of mine.

Adam: Thanks for dredging up that URL. That's an interesting set of comparisons and I largely agree with your assessment of the results. I'm slightly concerned about the accuracy of those tests, because the result they're showing for BlowUp doesn't look anything like what I've seen (except in pathological cases). Possibly it was an earlier version of the program, but it does raise a question in my mind.

By the way, I read your original message exactly as you intended. No misunderstandings or explanations or apologies needed at my end.

Scott: When I started this I was hoping for a simple, blanket answer. No such luck. It's all very, very conditional. I suspect a lot of the time I'm only going to find out the best way to enlarge a photo by doing it several different ways and comparing them. How annoying.

robert: I think you got taken in by some of the glowing reviews that I've seen, which implied gorgeous results enlarging photographs of a modest number of megapixels up to mural-sized. I don't know what those reviewers were photographing (or smoking!) but I've not seen much evidence to support that dream, not even in the sample photos on the software manufacturers' web sites. Wishful thinking.

As William so eloquently summed up (Great post, William!) it is difficult to create something from nothing. Theoretically, the program could be written that would do a much, much better job. The proof of that is that we can look at the photographs and determine, by purely visual analysis, what we would consider perceptually as noise, what we would consider signal, and what we would consider a real edge as opposed to an artifact). By this, I mean that we're not making the judgment based upon knowing what these objects looked like in the real world or what they are, we are just looking at the qualities of the image. That says that in principle a sufficiently smart program could do the same thing. But clearly it would have to be very, very smart. I think it's the kind of problem that you end up having to throw teraflops at.


pax / Ctein
[[ Please excuse any word-salad. ViaVoice in training! ]]
=========================================
-- Ctein's Online Gallery http://ctein.com
-- Digital Restorations http://photo-repair.com
=========================================

Dear Jeff,

Kind of off-topic but what the hell [grin].

If you have an Intel Mac running OS 10.4 or higher, you can forget about either of those options. ViaVoice is still sold but no longer under development, and it broke for me after OS X 10.3. iListen is now a dead product. Your sole option now is MacSpeech Dictate, which I have an order in for but hasn't shipped yet.

If you're running an older OS or a "G"- processor machine, you have other options including booting into Classic mode. That's what I'm doing all my dictation on at the present time, my old PowerBook running OS 9.

I have never been happy with voice recognition command-and-control. Even at 99 +% reliable, that other 1% of the time will send you off into never-never land if you're not paying attention. I tried it. After getting burned by several recognition errors, I found myself in the habit of issuing a voice command and then waiting to see if the machine executed successfully... and then continuing with my work. This was more disruptive and took longer than just using the mouse and keyboard!

If you're physically limited in a way that makes using a mouse and or a keyboard and or a digitizing tablet difficult, then you don't have any other choice. But if you have a choice, stick with the manual control methods and limit voice recognition to writing.


pax / Ctein
[[ Please excuse any word-salad. ViaVoice in training! ]]
=========================================
-- Ctein's Online Gallery http://ctein.com
-- Digital Restorations http://photo-repair.com
=========================================

My only experience with this level of upsizing was last year when I made a photo for a street adevertising, you can see it here:
http://www.albanogarcia.com.ar/imagenes/usina01.jpg
Nothing outstanding, and made for free, but it's great to see your pics huge on the streets.
They asked it in 1 x 1,5 meters at 300 dpi (3,5 by 5 feet), and I used my loved D50 with its 6mp and RAW. So I needed to upres huge. I tried Genuine Fractals, Bicubic in one step and Bicubic in several steps (10% increase each time). The best results were from Bicubic in several steps.
I'm amazed at how well the pic looks at this huge size, 6mp are versatile for sure. You can't see any pixel due to offset printing pattern when you get too close, and from a distance to see it complete, it looks great, with enough resolution.
Just my 2 cents...

The comments to this entry are closed.