There is a difference, yes. I am not sure if the difference is in the information in the data or in the resulting local contrast.
A quick test I did, showed that there was no visible difference in interpolation artifacts, for 100-200-100 compared to 100-120-100 up and down bicubic.
The only way to level the field then, would be to interpolate both images under comparison to a common larger size. Even if the interpolation affects local contrast, it would not remove detail from the higher resolving camera like dowsizing does.
Interpolate (verb)
- insert words into texts, often falsifying it thereby
Interpolate (verb)
- estimate the value of
Interpolate (verb)
- alter by insertion of new or spurious material
The FACT is, interpolation, by any method, adds information that was not in the original image, or it throws away information that was in the original image. At the same time, you can't fairly compare 100% views of images of different sensor resolutions. Despite anyone's assertions here, there is no right way to compare noise characteristics on screen. Each of the methods here represent a way to levelize the playing field, but each one of them contains flaws. Although I prefer Brainiac's method, there is a reason DPreview hasn't adopted it. DxO Mark has adopted even a different approach. I'm certain the testers at DPR, and DxO, and Brainiac all have understanding far beyond that which I have, but each have chosen a different method which reinforces my point. There is no "right" way to compare on screen.
Why is it better to interpolate only one image, and leave the other one unchanged, IF this change flaws the comparison? Isn't it better to transform both images with a process with equal artifacts?
kakomu wrote:
- Repeat the above step but resize the image back to its original resolution
The flaw in your steps is here. When you resize the image back to its original resolution you've added another variable to your test. You can't be sure if its the uprez or the downrez that causes your artifacts. You only know that doing both does.
brad_s wrote: timpdx wroteThis is why I still have my 5D even though I own a 7D
thx for posting.This is why I bought a 5 year old one.
All of this is why I tend to not care about ISO noise.
thedigitalbean wrote: kakomu wrote- Repeat the above step but resize the image back to its original resolutionThe flaw in your steps is here. When you resize the image back to its original resolution you've added another variable to your test. You can't be sure if its the uprez or the downrez that causes your artifacts. You only know that doing both does.
I know that the resize algorithm changes the image in both instances.
If you use nearest neighbor to resize the image up to a larger size and then back down to the original size, you'll find that at the resultant image is not different from the original image. The image after the process will result in a solid white canvas. Thus, we can assume that nearest neighbor doesn't add any information. (From what I've found on it, it seems to just enlarge data to fit the new space).
So, given this premise that we can resize up or down with nearest neighbor without adding much information, we can set up a pair of new tests. One where we upres with bicubic interpolation and downres with nearest neighbor and a second where we upres with nearest neighbor and downres with bicubic interpolation. Each of these tests will show whether bicubic resize filter changes the image.
In both tests, I've detected differences between the resultant image and the original image, indicating that the resize filter adds new information in both instances.
alundeb wrote:
Why is it better to interpolate only one image, and leave the other one unchanged, IF this change flaws the comparison? Isn't it better to transform both images with a process with equal artifacts?
In a good scientific test, you want to eliminate variability rather than add to it. I think that determining the overall noise of an image as a measure of the signal-to-noise ratio is the best method. This eliminates the transformation of interpolation and resize filters. It also gives you a measure of the entire frame rather than a small portion of the image. It also gives you a basis with which to compare noise to other sensors rather than to just a single sensor.
kakomu wrote: alundeb wroteWhy is it better to interpolate only one image, and leave the other one unchanged, IF this change flaws the comparison? Isn't it better to transform both images with a process with equal artifacts?In a good scientific test, you want to eliminate variability rather than add to it. I think that determining the overall noise of an image as a measure of the signal-to-noise ratio is the best method. From there, we can determine how resizing an image affects the signal-to-noise ratio.
This is not about numbers nor a scientific test. It is about a visual comparison that shows all information there is, among others noise level, noise character and image detail. Only noise level can be quantified with your flawed method. The interpolation method is the closest you get to printing at the same size.
Noise just one of camera parameter anyway, should it be fair to add contrast & resolution (sharpness), dynamic range, colour rendering into the comparison equation as well
alundeb wrote:
This is not about numbers nor a scientific test. It is about a visual comparison that shows all information there is, among others noise level, noise character and image detail. Only noise level can be quantified with your flawed method. The interpolation method is the closest you get to printing at the same size.
The only thing flawed here is trying to shoehorn a subjective analysis into an objective conclusion. A SnR measurement will determine noise level and image detail. Human eyes, on the other hand, are poor at determining comparative detail and noise level. For this reason, chroma subsampling is a very effective compression tool. However, human eyes are great at determining what you prefer.
Moroever, "noise character" is purely subjective.
An objective analysis of an image's noise will at least give a quantification of the noise. Otherwise, it's purely subjective discussion of which noise people prefer from which camera with pseudo-science to back it.
The interpolation method isn't close at all to seeing what you get when you print unless your monitor has a ppi size that is the same as your printer's dpi (or vice versa). Even scaling the image to show the same size on your monitor as your physical print (that is to say, if you print at A"xB", the image should take up A"xB" of physical space on your monitor) won't show the same thing.
If you want to compare prints, then print the images out and compare them (bring a loupe).
kakomu wrote:
The interpolation method isn't close at all to seeing what you get when you print unless your monitor has a ppi size that is the same as your printer's dpi (or vice versa). Even scaling the image to show the same size on your monitor as your physical print (that is to say, if you print at A"xB", the image should take up A"xB" of physical space on your monitor) won't show the same thing.
If you want to compare prints, then print the images out and compare them (bring a loupe).
A monitor has a fixed ppi, and the printer driver has a fixed ppi. This is totally independendt of the source image resolution. Your assertion about relative image size is plain wrong.
Edit: I misunderstood what you said here. Sorry.
Many people are not aware that Epson printer drivers for example, interpolate images using NN to 360 ppi (or 720 in some cases). This happens even if the print resolution is set to 1440 or 2880. For best results, one should interpolate the images to the native resolution using a better algorithm than NN before sending them to the printer. (Or use a RIP that bypasses the standard driver).
alundeb wrote:
A monitor has a fixed ppi, and the printer driver has a fixed ppi. This is totally independendt of the source image resolution. Your assertion about relative image size is plain wrong.
You've misunderstood my point. I know that the monitors have a fixed ppi. Though most printers I've dealt with allow you to change the dpi you print at (this is moot). The point I was trying to get across is that WYSIWYG does not apply here. The image you see on your screen will not look the same in print due to the fact that the screen and printer have two different image resolutions. Unless you print using the same dpi as your screen's ppi, the images will look different. This is on top of the fact that I've mentioned (many times) before that inkjet based printers produce non-discrete pixels based on the way that ink is absorbed into the paper.
alundeb wrote:
Many people are not aware that Epson printer drivers for example, interpolate images using NN to 360 ppi (or 720 in some cases). This happens even if the print resolution is set to 1440 or 2880. For best results, one should interpolate the images to the native resolution using a better algorithm than NN before sending them to the printer. (Or use a RIP that bypasses the standard driver).
Again, if you have to resize to a different resolution, this shows the affect of the resize filter on the image than it does anything else.
kakomu wrote: wickerprints wroteAnd you would be wrong.
You better read some texts on interpolation of data sets before you make such a blatantly incorrect assertion and embarrass yourself even further.
I was prepared to write a long treatise on the matter but realized that it would be futile. You are completely misunderstanding the nature of interpolation and the fact that there are other things going on with respect to the storage of the data that are not intrinsic to the interpolation algorithm.What am I wrong about?
The first image is the source image (black rectangle on white square). The second image is what happens to the (simple) image when you do a bicubic resize to a larger image (in this case from 400x400 to 1600x1600) and then a 2nd bicubic resize to the original resize. The edges are software. Edge fidelity is lost. Both images are saved as PNG, which is lossless, so any changes in the images are not caused by image compression.
You can easily reproduce the test in photoshop. The effect on more complex images is certainly going to be the same.
My point is that the quality of an image will change when you upres. The notion that upresizing doesn't "destroy" the image is clearly FALSE as interpolation will ultimately change the image. The only method of resizing an image to a larger size and back to a smaller size is to use nearest neighbor and only if you resize an image to a multiple of the original dimensions. Then again, who uses nearest neighbor for photo work?
Also, I'm not wrong as far as ink jets are concerned. THe ink does spread out once it hits the paper. This spreading and co-mingling of ink ultimately hides a lot of the digital resize artifacts that are present when you upres an image....Show more →
Only in this type of ultra simple artificial case is NN interpolation a better choice than bicubic or other schemes like Lancoz. In general for real world data, it is much worse and would be avoided at all cost.
kakomu wrote:
This is a fallacious argument. The only non-destructive method of resizing an image is nearest neighbor. This, of course, produces the worst results. As such, bicubic and bilinear resize algorithms are used frequently by those upresizing. Both of these algorithsm interpolate the image in way or another. Thus, the resize filter you use also determines the quality of the image and resize. Moreover, as I've already stated, by using the upresing method, you give an unfair advantage to cameras that have higher resolution as you magnify the noise when you upres.
You are dramatically overestimating the 'destructive' effect of uprezzing. I would challenge you to discern the difference between prints from any file and an uprez of it. You will do no better than a blind monkey. Pixels are not sacred, they are just vessels for data. The bit depth of our images is already at the subliminal level, so the quantising effect of uprezzing is essentially invisible at any print size where pixel borders can not be individually distinguished. The fact that you and many others don't trust basic arithmetic is your business, but I urge you to make even large prints from a file and its uprezzed copy where pixel boundaries are not resolved. You will not be able to see any difference.
kakomu wrote: SoundHound wrotePlease! Everyone try to understand what brainiac is saying otherwise it is vary hard to do these comparisons. Up rezzing doesn't destroy detail and you can't "throw away" data recorded by the bigger sensors.I understand what he's saying. I think his theory is flawed.
Moreover, upresing an image DOES destroy detail unless you're using nearest neighbor. The problem is that any resize algorithm, save for nearest neighbor, uses some form of interpolation. When you resize an image, you have to interpolate in some fashion. This interpolation cannot be undone. If you take an image, upres it and then resize it to the original size, the image will be different.
Please demonstrate that. I think you will be surprised by how conservative an uprez followed by a downrez is. Andi Dietrich illustrated this on a previous thread about this topic. Like I said, you are overestimating the purity of your pixels. Their values are only the result of an interpolation ('deBayer') by the time they reach your screen anyway. It's romantic, but wrong.
>It's easy to test if you use some discrete shapes. If make a black square on a white background, upres to a size that's not a multiple of the original resolution (using bicubic) and then downres the image to the original resolution, there are significant artifacts on the image.
Only at a scale that you do not wish to be resolved in prints.
Moreover, digital upresing of an image is NOT typical. As I've previously mentioned, printing an upresed image is not the same as viewing an upresed image on your screen. The print is analog and the co-mingled ink droplets make the resize artifacts difficult to discern.
...which is why you are even less likely to see the 'detrimental' effects of uprezzing in a print than on screen. You have made my point: any damage done is below the threshold of what a print will resolve, so ignore it and uprez at will.
It's probably worth pointing out that an image file is an attempt to record a continuum or field. Any notion of the purity of pixel values is misplaced reverence. Interpolated data is not random rubbish, it is carefully calculated data which is the best mathematical way to represent given data at a different sampling rate. A CD simply could not produce music if interpolation didn't work. Even the human eye has sampling rates. If you're uncomfortable with methods of discrete sampling to represent fields, and the manipulation of that data, you should get a different hobby.
Uprezzing is by far a more effective way of comparing two cameras than viewing at different magnifications. The latter method gives errors many orders of magnitude greater than any sampling noise introduced by bicubic uprezzing. If you don't like the loss of sharpness (not detail) at 100%, then add USM to taste.
Since we're on the subject of how bicubic uprezzing turns a picture of a unicorn into a picture of a pantomime dromedary, one more thing before I go to bed: a web site wishing to overcome these silly objections to uprezzing could do all three of the following.
1) compare an exact nearest neighbour quadrupling of pixels in the lower rez file to a bicubic uprez of the higher rez file, both images having the same number of pixels.
2) compare a bicubic uprez of the lower rez file to the original higher rez file, both images having the same number of pixels.
3) compare a bicubically downrezzed higher rez file to the lower rez file, both images having the same number of pixels.
In each case I would propose that the bicubically interpolated file could be USM'ed to match the other. Method 3 will be the least accurate, but good enough for most people's purposes. The other two methods shouldn't disagree significantly. The benefit of posting all three methods is that those who don't understand why method 2 is perfectly reasonable won't be tempted to proliferate their bogus meme.
There is a 4th method, namely uprezzing both files bicubically to some size larger than each, but it's time consuming and pretty pointless.
Pixel Perfect wrote:
Only in this type of ultra simple artificial case is NN interpolation a better choice than bicubic or other schemes like Lancoz. In general for real world data, it is much worse and would be avoided at all cost.
I understand, which is why I also wrote that NN produces "the worst results".
brainiac wrote:
Uprezzing is by far a more effective way of comparing two cameras than viewing at different magnifications.
You cannot come to any conclusion based on upresing an image. Or, any conclusion you do come to will be dubious at best. As I've already mentioned before, upresing an image is outside of what people normally do beyond printing. And printing produces an image that far different looking than what you see on the screen.
kakomu wrote: Pixel Perfect wroteOnly in this type of ultra simple artificial case is NN interpolation a better choice than bicubic or other schemes like Lancoz. In general for real world data, it is much worse and would be avoided at all cost.I understand, which is why I also wrote that NN produces "the worst results".
brainiac wroteUprezzing is by far a more effective way of comparing two cameras than viewing at different magnifications.You cannot come to any conclusion based on upresing an image. Or, any conclusion you do come to will be dubious at best. As I've already mentioned before, upresing an image is outside of what people normally do beyond printing. And printing produces an image that far different looking than what you see on the screen.
Just to clarify, you are arguing that half of the degradation in the following before/after example is making it impossible to compare two images shot with different cameras: http://cyberphotographer.com/1d4/beforeafter.jpg
I hardly see any degradation of the image here, and certainly not enough to influence a comparison between two cameras. Remember, this is the result of two consecutive interpolations, not one. A modest amount of sub-pixel USM would mitigate any barely visible softening still further. Thanks for the example - it's a win for common sense and maths, and a loss for superstition. Your concern with the detrimental effects of uprezzing is entirely academic.