Reading mode: Light Dark

resolution tests 5MPvs.12MPyeilds unexpected results

Go to last post
Leon Wittwer said:
If the 13 MP Bayer image pixel size is x, then the resolving power
would be approximately 2.2x, that is, two lines about 2.2x apart
could be discerned (MTF~0.5). Assuming a very good down rezzing
algorithm, you could down rez until the new pixel size is
approximately 2.2x without loosing significant resolving power.
That is about a 2.7 MP size. Thus, I would expect virtually no
difference at 8 MP or at 5 MP in the resolving power and print
quality at virtually any sized print even assuming some resolving
power loss from the down rezzing algorithm.
Click to expand...
This is exactly right. In fact, on the Olympus forum I ran two tests where I took 5 8Mp E-300 images of different subject matter and used in-camera processing to produce 8Mp SHQ and 5Mp SQ images. I then uprezzed the 5Mp SQ images back to 8Mp in Photoshop 7 using bicubic interpolation. I then invited forum members to look at side-by-side 100% crops and try to ID which image was which. The results were on average about 50% right. In other words, most people could not tell which was better. There were, in fact, small subtle differences and some people were able to learn from the first trial and use what they learned from those results to improve their results on the second trial. But even with that advantage, the overall result was essentially a wash.

So you are correct. A downsized 5Mp image from an 8Mp CFA image contains essentially the same amount of image detail as the 8Mp CFA image. There is a lot of spatial "fluff" in a native sized CFA image.

So there should have been almost no visual difference between the 12Mp and the 8Mp image and there should have been some visual difference between the 5Mp image and the other two. But if the subject wasn't demanding, then you could easily miss that difference.

http://www.jayandwanda.com/dpreview/sq2/

http://forums.dpreview.com/forums/read.asp?forum=1022&message=20651503

http://forums.dpreview.com/forums/read.asp?forum=1022&message=20696691

--
Jay Turberville
http://www.jayandwanda.com
 
Jay Turberville said:
So you are correct. A downsized 5Mp image from an 8Mp CFA image
contains essentially the same amount of image detail as the 8Mp CFA
image. There is a lot of spatial "fluff" in a native sized CFA
image.
Click to expand...
A lot of this 'fluff' is going to be down to the 'fuzzing' (gee, I love these technical terms...) of the AA filter.

Has anyone worked out a typical number that relates how much less fuzzy a picture taken with a camera without an AA filter is compared to one with (I realise the strength of the filter will be variable though with different cameras). Thus could one say that say a 10MP D200 in fuzzed condition would give the same sharpness as say an 8MP defuzzed one (if one existed)

Ah ... these thoughts that keep us from going to bed ......
 
Big Ga said:
A lot of this 'fluff' is going to be down to the 'fuzzing' (gee, I
love these technical terms...) of the AA filter.

Has anyone worked out a typical number that relates how much less
fuzzy a picture taken with a camera without an AA filter is
compared to one with (I realise the strength of the filter will be
variable though with different cameras). Thus could one say that
say a 10MP D200 in fuzzed condition would give the same sharpness
as say an 8MP defuzzed one (if one existed)
Click to expand...
I'd assume that the answer is subject dependant. That is why some demosaicing algorithms and settings work more or less well on different subjects. The number I use to estimate the "fluff" is 80% of the linear resolution implied by the pixel count and is derived from a survey of resolution chart results on DPreview.

if the AA filter is less agressive, then the odds of getting demosaicing errors increases. As it is, most AA filters let the camera still produce some mild color moire.
Big Ga said:
Ah ... these thoughts that keep us from going to bed ......
Click to expand...
Yep. This is one of the reasons I'm actually a fan of megapixels. Its isn't so much to get more final resolution, but to get results that are less compromised by a low pass filter or at least to be able to use a less agressive low pass filter.

My bet is that the genius who did this comparison test isn't up at night thinking about this stuff though.

--
Jay Turberville
http://www.jayandwanda.com
 
This argument sounds specious to me. Imagine a 12 Mp sensor (e.g. 4000 x 3000 pixels) of a hypothetical camera with an objective that has infinite resolving power. Suppose then we take an image of 3000 horizontal, alternate black and white lines such that each line is 1 by 4000 pixels. We thus have an image with 1500 black lines interlaced with 1500 white lines.

Now what happens when we reduce to 5 Mp (say 2500x2000)? Here we could only have an image with 1000 black lines interlaced with 1000 white lines. In addition our 4000 pixel lines would become 2500 pixel lines.

We could never get an image of 3000 horizontal lines in a format that had less than 3000 pixels in its vertical dimension. Information (detail) will therefore always be lost and no algorithm can avoid that.

A different question is whether this level of detail matters and that's just a question of how big the pixels are - i.e. can the human eye distinguish?

FrankyM
 
You are completly right, and it is this inability to resolve details that results in moire.
 
FrankyM said:
This argument sounds specious to me. Imagine a 12 Mp sensor (e.g.
4000 x 3000 pixels) of a hypothetical camera with an objective that
has infinite resolving power. Suppose then we take an image of 3000
horizontal, alternate black and white lines such that each line is
1 by 4000 pixels. We thus have an image with 1500 black lines
interlaced with 1500 white lines.
Click to expand...
It's not as simple as that. Firstly, the only cameras that would be theoretically capable of resolving individual white-black line pairs would be those with Foveon sensors. And even then, you could only do this if you managed to align the sensor pixels perfectly with the lines, and as you said, used an ideal lens.

The reason is that bayer sensors require an anti aliasing filter, which blurs the image. If for example your camera has a bad demosaicking algorithm, you might find that virtually all the detail delivered at full-size output could be held in an image with a quarter of the MP's. With a better processing system in camera, it'll be between half and two thirds.

To go back to the original debate - 5 vs 8 vs 13 - Try the following:

Download the test scene shot from the 5D review, and then download one from a 5mp camera. Down-res the 5D shot to 5mp's, then up-res it back to 13, and if you like, apply a gentle USM. Do the same from 13 to 8 and back. Now up-res the shot from the 5mp cam to 13mp's, and compare detail (this is like comparing large prints). You might be surprised at the difference.
FrankyM said:
A different question is whether this level of detail matters and
that's just a question of how big the pixels are - i.e. can the
human eye distinguish?
Click to expand...
Now you're talking sense :-)

It also depends on subject matter - a shallow DOF portrait simply doesn't require as much detail as a landscape with bushes, rocks, trees and distant mountains.

Cheers,
Andrew
 
Yes it is as simple as that. Do a simple test. Take a 6Mp RGB file (Raw, JPG, whatever you want). Note the size in bytes. Reduce to 3 Mp, save in same format, same quality level if JPG. Note the size of that.

The difference in bytes is the information difference, and that translates to detail in the image.

FrankyM
 
FrankyM said:
Yes it is as simple as that. Do a simple test. Take a 6Mp RGB file
(Raw, JPG, whatever you want). Note the size in bytes. Reduce to 3
Mp, save in same format, same quality level if JPG. Note the size
of that.

The difference in bytes is the information difference, and that
translates to detail in the image.
Click to expand...
Nope. A file can have more bytes than information. You can prove it simply enough.

Take any image file and uprez it to twice its original size. Now, does it have more information? OK, now downrez it back to the original. Isn't the information about the same as in the original (minus losses due to interpolation algorithm issues)?

So, it is clear that files size does not necessarily equate to image detail or information content.

This applies in this case because images from CFA sensors are not only softened slightly by a low pass filter, but the are essentially "uprezzed" as well. And as I outlined in a previous post, it is very difficult for most people to tell the difference between an 8Mp image direct form a CFA camera and an 8Mp image that has been downrezzed to 5Mp and then uprezzed back to 8Mp.

--
Jay Turberville
http://www.jayandwanda.com
 
Yes a file can (and does in all cases) have more bytes than simply required to define the image. Nevertheless, what is reduced in a file when we down rez is pixel information and that, sir, is detail.

I repeat. If you down rez an image, you lose detail, full stop. You may not be able to notice on a print of some particular size but that is another story.

FrankM
 
FrankyM said:
Yes a file can (and does in all cases) have more bytes than simply
required to define the image. Nevertheless, what is reduced in a
file when we down rez is pixel information and that, sir, is detail.
Click to expand...
For continuous tone images, every resampling algorighm wil introduce some degree or error or loss. In that sense, any downrezzing of an image does lose some information. But the practical reality is that it loses very little information - especially in proportion to the actual loss in numbe of bytes needed to record the image. We can reduce a CFA image in size by over 35% and lose almost no information - not the 35% that some people would assume.
FrankyM said:
I repeat. If you down rez an image, you lose detail, full stop. You
may not be able to notice on a print of some particular size but
that is another story.
Click to expand...
Just about any time you process an image - especially if you resample it, you lose detail. But the simple fact is that you can reduce a typical native resolution CFA image by 35% of its original image size and suffer almost zero loss in detail or information.

The point is that an image downsized to 8Mp from a 12-13Mp CFA sensor will have almost the same amount of picture information as the non-downsized image and a significantly greater amount of detail than a native sized 8Mp image directly from an 8Mp camera. Very little is lost.

--
Jay Turberville
http://www.jayandwanda.com
 
Warren Sarle said:
Warren Sarle said:
Firstly, the only cameras that would be
theoretically capable of resolving individual white-black line
pairs would be those with Foveon sensors.
Click to expand...
Nonsense.
Click to expand...
I think he meant where the line spacing is equal to the pixel spacing. I haven't seen any test results that managed this without introducing color errors.
Warren Sarle said:
Warren Sarle said:
The reason is that bayer sensors require an anti aliasing filter,
which blurs the image.
Click to expand...
The Leica M8 and Kodak SLR/n/c do not have AA filters.
Click to expand...
And the Kodaks exhibit a color artifacts at sharp high contrast edges. It will be interesting to see how the M8 does. Its been a few years since a camera maker decided to put out a camera without one. Maybe better demosaicing algorithms will make reduce the problem seen with the Kodaks.

--
Jay Turberville
http://www.jayandwanda.com
 
Jay Turberville said:
Jay Turberville said:
Jay Turberville said:
Firstly, the only cameras that would be
theoretically capable of resolving individual white-black line
pairs would be those with Foveon sensors.
Click to expand...
Nonsense.
Click to expand...
I think he meant where the line spacing is equal to the pixel
spacing. I haven't seen any test results that managed this without
introducing color errors.
Click to expand...
That's right, I was referring to lines exactly one pixel wide, i.e. line pair exactly two pixels wide.
Jay Turberville said:
Jay Turberville said:
Jay Turberville said:
The reason is that bayer sensors require an anti aliasing filter,
which blurs the image.
Click to expand...
The Leica M8 and Kodak SLR/n/c do not have AA filters.
Click to expand...
Click to expand...
True. I'm interested to see how Leica handles this. Take a look at Kodak's attempt:
http://www.dpreview.com/reviews/kodakslrc/page24.asp

Andrew
 
FrankyM said:
Yes a file can (and does in all cases) have more bytes than simply
required to define the image. Nevertheless, what is reduced in a
file when we down rez is pixel information and that, sir, is detail.

I repeat. If you down rez an image, you lose detail, full stop.
Click to expand...
Yes, of course you lose some detail. But when you're dealing with a typical digital camera's native output it is not as much as people think, and when the down-sampled image is compared to a full size image taken with a lower resolution camera, the down-sampled image will usually look significantly better. Try it, as I explained in my earlier post, and let your eyes be the judge.
FrankyM said:
You may not be able to notice on a print of some particular size but
that is another story.
Click to expand...
Well actually that's where this whole discussion started.

Don't simply judge by number of bytes. I can make a 300 MB file out of the dpreview logo at the top of the screen if I really want to.

Andrew
 
AndrewJWH said:
True. I'm interested to see how Leica handles this. Take a look at
Kodak's attempt:
http://www.dpreview.com/reviews/kodakslrc/page24.asp
Click to expand...
No sane person would deny that the lack of AA filter does bring its own problems in certain situations.

However no sane photographer is going to shoot nothing but line pair diagrams. I couldn't tell you the actual average statistics, but I'd be prepared to wager a not insubstantial burger that for every one photograph taken with a non AA filtered Kodak that does indeed have troublesome moire patterns, there are dozens/hundreds (maybe thousands for some subject matter) of shots that do NOT have any show stopping artifacts and make your normal AA filtered camera pics look soft and mushy by comparison.
 
I do not judge by the number of bytes, nor do I state anywhere that the loss of detail is significant. Indeed I clearly state that losing detail is one thing, but being able to see the difference is another.

You state that down rezzing to 5 Mp from 12 Mp will (depending on the algorithm used) will make a 5 Mp file different (and therefore a different image) to that produced by a 5 Mp camera. Nevertheless it will not contain more detail.

If you go back to the original article, i.e. "Resolution tests 5MPvs.12MP yields unexpected results" what is being stated is that no-one could tell the difference between the prints obtained using the 12Mp image and a 5 Mp downrezzed version.

The person to whose input I replied inferred/stated that detail is not lost down rezzing to 5Mp so the test is flawed and should be compared to a 5 Mp camera.

I say that that argument is flawed and that the 5Mp downrezzed image contains no more detail than a 5 Mp image from a camera.

FrankyM
 
FrankyM said:
I say that that argument is flawed and that the 5Mp downrezzed
image contains no more detail than a 5 Mp image from a camera.
Click to expand...
Exactly. 5million pixels is 5million pixels.

Some people need to get sense of a cogent arguement, intend of making wild claims that don't stand up. Such as this iditioc gem
FrankyM said:
Firstly, the only cameras that would be
theoretically capable of resolving individual white-black line
pairs would be those with Foveon sensors.
Click to expand...
Not only is that nonsense, if you really wanted to get technical about it, the foveon sensor will actually hamper that ability.

Instead trying to get all clever and technical and wasting time on this forum, I would suggest people go out take a well detailed shot with a 5mp and then with a 12mp camera, post process them as you would normally.print them both at 16x24, as the OP and compare the difference.

There will be one.
 
riddell said:
riddell said:
I say that that argument is flawed and that the 5Mp downrezzed
image contains no more detail than a 5 Mp image from a camera.
Click to expand...
Exactly. 5million pixels is 5million pixels.
Click to expand...
See my earlier post, and try the experiment so you can see for yourself.
riddell said:
Some people need to get sense of a cogent arguement, intend of
making wild claims that don't stand up. Such as this iditioc gem
Click to expand...
True. However, perhaps it would be best to clean up the spelling and grammar issues first, and then talk about cogency.
riddell said:
riddell said:
Firstly, the only cameras that would be
theoretically capable of resolving individual white-black line
pairs would be those with Foveon sensors.
Click to expand...
Not only is that nonsense, if you really wanted to get technical
about it, the foveon sensor will actually hamper that ability.
Click to expand...
Why will the Foveon hamper that ability? Granted, it will show aliasing as soon as detail delivered by the lens passes Nyquist frequency for the sensor, but up to that point it's pretty good. That's why people claim that the 3.4MP file the Sigma puts out is basically as good as a file from a typical 6MP bayer senor camera.
riddell said:
Instead trying to get all clever and technical and wasting time on
this forum, I would suggest people go out take a well detailed shot
with a 5mp and then with a 12mp camera, post process them as you
would normally.print them both at 16x24, as the OP and compare the
difference.
Click to expand...
Of course there'll be a difference - a big one in my opinion, but this is not a new concept.
Now, not to flog a dead horse or anything, but quoting the OP:

"Same exact photo, down-rezzed twice, all three printed at the same poster size."

Cheers,

Andrew
 
Leon Wittwer said:
If you shoot landscapes and natural things, then anti-aliasing
filters are not quite as necessary.
Click to expand...
Er.... I would say that HAVING an AA filter can be a distinct DISADVANTAGE!
Leon Wittwer said:
If you photography people and
man made things where there may be fine near periodic detail, then
such filters are necessary.
Click to expand...
I believe I did say
"lack of AA filter does bring its own problems in certain situations"

However, just to be pedantic, to say that AA filters are 'necessary' if you photograph "people and man made things", implies that you cannot shoot these types of things if your camera doesn't have an AA filter. That is not correct.
Leon Wittwer said:
Certain fabric patterns, for example,
cause significant problems.
Click to expand...
I completely agree.

My limited experience (but fairly extensive research) would however lead me to believe that in many circumstances, a suitable RAW converter can significantly reduce problems in PP.
 
You must log in or register to reply here.

Keyboard shortcuts

f
Forum
m
My threads
Mobile site
About
Editorial content
Cameras & Lenses
Community
All content, design, and layout are Copyright © 1998–2025 Digital Photography Review All Rights Reserved.
Reproduction in whole or part in any form or medium without specific written permission is prohibited.
When you use DPReview links to buy products, the site may earn a commission.
©GPS Media - Guides, Products, Services.
Back
Top