fredmiranda.com Login
This topic is in the 2010 archive. You can read it and like its posts here. It is read only.

Archive 2010Noise, Interpolation, and Comparison 101

  
 
thedigitalbean
Offline
Details
p.7 #1Noise, Interpolation, and Comparison 101
kakomu wrote:
Moreover, the method used to demosaic an image is irrelevant, so long as it's consistently used to test images. It's a necessary evil that cannot be circumvented as far as digital photography is concerned, save for sensors that are similar to film emulsions with several transparent layers that are, each, reactive to different colors. However, in the end, the final image that we get from the raw converter is what's important to the testing.

Though, simply because an image is demosaiced does not mean that it's plausible to degrading an image further with other interpolating algorithms. Rather, all it will show
...Show more →

You claim that interpolation done when resizing up degrades the image. I'm sure there will be others that agree with you. I and several others think that claim is wrong and without merit. As far as I can tell we are at an impasse here and it may be simply better to leave it at "agree to disagree".
Feb 04, 2010 at 11:29 AM
brainiac
Offline
[X]
Details
p.7 #2Noise, Interpolation, and Comparison 101
kakomu wrote:
The point is that upresing an image exaggerates the noise in an image.


No.

Comparing noise in 10% of one image to 15% of another image exaggerates the noise in the 10% sample. Visibility of noise is proportional to magnification. That is why uprezzing is necessary to avoid the simple error of unequal magnification.

Edited on Feb 05, 2010 at 05:37 AM · View previous versions
Feb 05, 2010 at 05:10 AM
brainiac
Offline
[X]
Details
p.7 #3Noise, Interpolation, and Comparison 101
skibum5 wrote:
so just what the heck is this 'pantomime' dromedary camel of yours


A pantomime dromedary is a camel costume worn by two people, one forming the head and front legs, and the second bending over to form the torso and rear legs. The effect is unconvincing and comic, especially when the two actors disagree about the dromedary's direction. If you saw a unicorn, and photographed it, you would be surprised and disappointed to find instead a pantomime dromedary in your pictures.

I have been looking for significant image degradation through bicubic uprezzing for about two decades, and it has still never happened. Sometimes I am tempted to add a little USM to get a better match. I can understand that some people develop irrational sentimental attachments to the beautiful squareness of their pixels, but they should realise that those squares were never really there in the original scene, and their utility lies entirely in describing the field of light around which they are only a malleable vessel. The squareness of those pixels only affects the sharpness when printing at or near the resolution limit of the file, in which case addition of USM, and even sub-sensel noise, will compensate the rounding of pixel edges.
Feb 05, 2010 at 05:27 AM
brainiac
Offline
[X]
Details
p.7 #4Noise, Interpolation, and Comparison 101
32067dlm wrote:
So, how does the 1D Mark IV compare to itself?


It is much noisier than itself, and much less noisy than itself, if you compare it to itself at different magnifications. Viewing a raw file and an sraw file at 100% will illustrate this perfectly.
Feb 05, 2010 at 05:40 AM
Mike Mahoney
Offline
Details
p.7 #5Noise, Interpolation, and Comparison 101
BigBlueDodge wrote:
I think one obvious point we are all missing here is to appreciate what a damn fine camera the 5D classic was and still is. It's impressive that almost 5 yr old technology is still hanging in there with the new kids on the block.


It's the pixel pitch (size) .. the 5D has larger pixels which inherently have less noise and wider dynamic range.
Feb 05, 2010 at 06:52 AM
jkurkjia
Offline
Details
p.7 #6Noise, Interpolation, and Comparison 101
thedigitalbean wrote:
You claim that interpolation done when resizing up degrades the image. I'm sure there will be others that agree with you. I and several others think that claim is wrong and without merit.


Much of this discussion depends on your definition of "degrade." Interpolation algorithms create subjectively objectionable artifacts (depending on the type of data and its associated spatial frequency). It's not a stretch to refer to any objectionable artifact output from an interpolation process as a degradation (though it's relative in the case of Interpolation Scheme A versus B).

If resizing up is such a straightforward, pure, and perfect action then why are there more interpolation schemes available than you can shake a stick at?

I've said it before and will repeat it once again, for a screen comparison of two picture files having different resolution both files should be interpolated (one down and the other up) to an average size; this approach will somewhat equally contaminate both files with artifacts associated with whatever interpolation algorithm is chosen. It should be noted that choosing an algorithm isn't exactly like shooting fish in a barrel; some are better going up in size while others excel at going down.

Note, on the subject of "comparison of two picture files having different resolution" I could easily live with "only resizing up" if, as part of the comparison both files were resized up to something like three or four times the average size of the two files.

Regards,

Joe Kurkjian
Feb 05, 2010 at 06:56 AM
keithreeder
Offline
Details
p.7 #7Noise, Interpolation, and Comparison 101
brainiac wrote:
some people develop irrational sentimental attachments to the beautiful squareness of their pixels, but they should realise that those squares were never really there in the original scene, and their utility lies entirely in describing the field of light around which they are only a malleable vessel.


That's downright poetic, Richard!


Feb 05, 2010 at 07:15 AM
chez
Offline
Details
p.7 #8Noise, Interpolation, and Comparison 101
Dawei Ye wrote:
Interestingly who actually prints nowadays? I don't think I've printed an image in 12 months I find most clients just want digital files nowadays for Facebook or emailing to family or whatever


Then I say it matters very little which recent DSLR you use if all you do with them is look at the images online. All this measurebating is just that, measurebating.
Feb 05, 2010 at 07:16 AM
chez
Offline
Details
p.7 #9Noise, Interpolation, and Comparison 101
alundeb wrote:
I don't get it either how upresing could be detrimental to a comparison. The argument that upresing changes the pixel data falls on itself, because pixel data are always changed before any normal presentation: web, projector, TV screen, print, etc. If the "holy unchangeable pixel" argument was vaild, it would be impossible to compare the quality of any final output, inluding prints and web resized images.


Uprezing and viewing at 100% on the screen gives you 1 view. However, printing at the paper size you want will give you a totally different view. The print is much more foregiving than an image viewed at 100%. It also really depends on the paper being printed on.

To me, all this comparing at the pixel level is truly a waste of time. If the image is destined to be viewed on the screen, who here sends their photos onto something like flikr at 100% magnification?

If the image is to be printed, then there are so many other variables that affect the final print that looking at individual pixels and comparing is just a waste of time.

Whatever happened to the photo part of photography.
Feb 05, 2010 at 07:22 AM
chez
Offline
Details
p.7 #10Noise, Interpolation, and Comparison 101
kakomu wrote:
alundeb wroteI don't get it either how upresing could be detrimental to a comparison. The argument that upresing changes the pixel data falls on itself, because pixel data are always changed before any normal presentation: web, projector, TV screen, print, etc. If the "holy unchangeable pixel" argument was vaild, it would be impossible to compare the quality of any final output, inluding prints and web resized images.
The point is that upresing an image exaggerates the noise in an image. Exaggerating a defect is dishonest.

And, as I've stated numerous times, people don't typically upres an image with the intent of keeping it in a digital form (unless they plan on heavily manipulating it). We all know that upresing an image makes it look worse than having an intrinsically higher resolution image. This is a no-brainer. Thus there are multiple reasons why upresing an image will always make an image look worse than one with a higher resolution sensor.

The problem is that the conclusions drawn from this exaggeration
...Show more →

Well said and I could not agree more. Print the image onto the paper of choice if you really want to know how the two images would compare. Doing anything else is wasting your time.
Feb 05, 2010 at 07:25 AM
chez
Offline
Details
p.7 #11Noise, Interpolation, and Comparison 101
alundeb wrote:
kakomu wroteThus, we come full circle: The only reason to upres an image is to print the image. Printed images will look different than digital images on-screen. Thus, any conclusion about an upresed image on-screen DOES NOT apply to the actual printed image. Thus, making on-screen comparisons with an upresed image HAS NO MERIT.
When printing, there is always a relationship between the quality of the file and the quality of the print. The fact that the print looks different, does not remove the fact that the quality difference between prints is proportional to the quality diffrence between upresed files. Inspection of upresed files is very important to acheive a large print with fine detail.
For smaller output, it is still essential to compare at the same magnification.

In addition, a 100% crop is already interpolated by the RAW converter, so there is no such thing as visual comparison of uninterpolated images.


I agree to a degree. There is a relationship between the 100% and the quality of the final print, but it is not a direct relationship. The 100% crop definetly exaggerates this relationship and from my experience ( I print photos up to 24x30 on a regular basis ), the final print is much more foregiving than the 100% crop.
Feb 05, 2010 at 07:31 AM
alundeb
Offline
Details
p.7 #12Noise, Interpolation, and Comparison 101
chez wrote:
alundeb wrote
kakomu wroteThus, we come full circle: The only reason to upres an image is to print the image. Printed images will look different than digital images on-screen. Thus, any conclusion about an upresed image on-screen DOES NOT apply to the actual printed image. Thus, making on-screen comparisons with an upresed image HAS NO MERIT.
When printing, there is always a relationship between the quality of the file and the quality of the print. The fact that the print looks different, does not remove the fact that the quality difference between prints is proportional to the quality diffrence between upresed files. Inspection of upresed files is very important to acheive a large print with fine detail.
For smaller output, it is still essential to compare at the same magnification.

In addition, a 100% crop is already interpolated by the RAW converter, so there is no such thing as visual comparison of uninterpolated images.
I agree to a degree. There is a relationship between the 100% and the quality of the final print, but it is not a direct relationship. The 100% crop definetly exaggerates this relationship and from my experience ( I print photos up to 24x30 on a regular basis ), the final print is much more foregiving than the 100% crop.


The relationship depends on the resolution and how you print. For the best quality, I resize to 720 ppi and print with "Finest Detail" enabled. The image resolution then matches the highest possible printer driver resolution. When the original image has a resolution at about 180 ppi or lower, the printer is very little forgiving. I can see everything that was in the original file on the print. From there, I can see improvement in quality all the way up to 480 ppi original file, but the improvement diminishes gradually.
So to my eyes, the relationship is close to linear at original resolutions below 180 ppi and the quality saturates gradually up to about 480 ppi.
Feb 05, 2010 at 07:53 AM
chez
Offline
Details
p.7 #13Noise, Interpolation, and Comparison 101
alundeb wrote:
chez wrote
alundeb wrote
kakomu wroteThus, we come full circle: The only reason to upres an image is to print the image. Printed images will look different than digital images on-screen. Thus, any conclusion about an upresed image on-screen DOES NOT apply to the actual printed image. Thus, making on-screen comparisons with an upresed image HAS NO MERIT.
When printing, there is always a relationship between the quality of the file and the quality of the print. The fact that the print looks different, does not remove the fact that the quality difference between prints is proportional to the quality diffrence between upresed files. Inspection of upresed files is very important to acheive a large print with fine detail.
For smaller output, it is still essential to compare at the same magnification.

In addition, a 100% crop is already interpolated by the RAW converter, so there is no such thing as visual comparison of uninterpolated images.
I agree to a degree. There is a relationship between the 100% and the quality of the final print, but it is not a direct relationship. The 100% crop definetly exaggerates this relationship and from my experience ( I print photos up to 24x30 on a regular basis ), the final print is much more foregiving than the 100% crop.
The relationship depends on the resolution and how you print. For the best quality, I resize to 720 ppi and print with "Finest Detail" enabled. The image resolution then matches the highest possible printer driver resolution. When the original image has a resolution at about 180 ppi or lower, the printer is very little forgiving. I can see everything that was in the original file on the print. From there, I can see improvement in quality all the way up to 480 ppi original file, but the improvement diminishes gradually.
So to my eyes, the relationship is close to linear at
...Show more →

Depends on what paper you print. I print on both gloss ( Epson Exhibition Fibre ) and canvas and an image printed on the gloss might show some degradation whereas the same image printed on the canvas is fantastic.
Feb 05, 2010 at 08:50 AM
brainiac
Offline
[X]
Details
p.7 #14Noise, Interpolation, and Comparison 101
Unlike you fellows, I choose my cameras with deliberate disregard for specific existing output technologies. I take my pictures for posterity, and I have no idea at what size or resolution any picture may be used. I don't even know if someone will wish to crop it.

Comparison on a good monitor is a perfectly good, albeit imperfect way to assess a pair of cameras' OUTPUT DEVICE-INDEPENDENT PERFORMANCE. Looking at different monitors and various print types helps too, but by and large a good monitor can show you a lot about the difference between two cameras. Certainly the differences may be marginal, but as I have said, if you are going to shoot hundreds of thousands of other people's once-in-a-lifetime pictures on your cameras, then you have not just the right, but a duty, to be fussy about camera output.

All this stuff about how "you can't judge the differences between cameras on screen" is utter balls. But in order to make that judgment, you must be careful to ensure that you are judging fairly, and that means at equal magnification.

Joe's 'meet in the middle' suggestion is perfectly fair, except for two flies in the ointment. First, the higher rez file has been handicapped in a way that the lower rez file hasn't. A 16 Mpixel file can hold all the data found in a 10 Mpixel file, but it can't hold all the data found in a data-saturated 21 Mpixel file. Secondly, the points about interpolation algorithms ignore the fact that different algorithms favour different subject matter. For instance, the nearest neighbour method would have done a beautiful job of uprezzing and downrezzing the black-on-white box in the sample earlier in the thread - there would have been zero softening. The same goes for other algorithms. Some favour patterns, some work well with skin, some are good for buildings. In general, bicubic, followed by optional sharpening does a perfectly adequate job for the sake of uprezzing a lower rez file for comparison with a higher. Allowing an optional USM after bicubic uprezzing should visually satisfy the most pedantic of testers, since she would not likely be able to reliably distinguish the files at any useful output resolution.

Uprezzing both files to a much higher resolution certainly answers many of the over-fussy criticisms of the convenient method I suggest, but it is time consuming, and the results are ultimately indistinguishable from the simple method I always propose.

Any method involving downrezzing is OK to see if a 21 Mpixel camera is as good as a 12 Mpixel camera, but it's often not explicit about whether the 21 Mpixel camera is actually better. So don't downrez, as that won't tell you the whole story. Either uprez both files to some massive size, or if you recognise that it will amount to practically the same thing, do as I have always said, and uprez the lower to the higher. It's a great way of comparing cameras, and is often better than doing so behind the smokescreen of a printer driver, ink technology, blotting paper, and limited size print.
Feb 05, 2010 at 01:23 PM
FM ForumsCanon ForumJoin Upload & Sell
  
 

You are not logged inLog inRegister