Let's finally resolve the D800 downsampling / NR questions

[barnicle]

I can't wait for a month from now when we can see the images and be done with the theoretical debate.

 

[wombat779]

wombat779 wrote:
Hey wombat, sorry I didn't got back to that thread I guess.
Here is what I saw:

Thanks Dan. Looking at your sample, I do see somewhat less pixel-level noise in the 16MP bicubic sample than in the 36MP sample. This is consistent with what I saw running the test. If I take the test further and keep reducing, this is what I get:





To my eye, at each reduction the pixel level noise becomes less apparent, and by the lowest reduction, it is almost clean. Of course, at that level the image itself is quite small, but I think demonstrates that simple downsampling without other noise reduction acts to reduce apparent pixel-level noise.

 

[Horshack]

This looks very promising, once I fully understand the experiment ;-)

You're using 3x FL and light ratios to simulate a high-MP vs 1/3 MP sensor, in terms of light received, acuity (including diffraction effects), correct?

The different image scales in your 1-6 images is throwing me of the scent a bit, even though I know those scales are necessary for the experiment. Please tell me if these statements are correct

• Image #2 depicts the low-MP sensor at its native resolution?
• Image #3 depicts the high-MP sensor at its downsampled (1/3) resolution?
• Image #4 is a shadow noise comparison at low-MP native vs high-MP downsampled?
• Image #5 and #6, I'm confused about what scale is being compared. Is it native low-MP vs native high-MP?

 

[The_Suede]

Yes on all accounts, but with addition:

the downsampled 150mm images are downsampled after a "normal" but weak noise reduction. They're not "straight" downsamples.

 

[danharperphotography.com]

To my eye, at each reduction the pixel level noise becomes less apparent, and by the lowest reduction, it is almost clean. Of course, at that level the image itself is quite small, but I think demonstrates that simple downsampling without other noise reduction acts to reduce apparent pixel-level noise.

Ok I see where you're coming from and you're totally right, reducing the image, reduces the visible noise, I'll give you that. But you had to reduce it to an unusable size to make it work to your point. The problem with that is just like I said, if you reduce your dinner to just one slice of steak and tell your stomach that you're reducing the meal for diet purposes, your stomach will punch you in the face (no seriously it will! ) and you will still be hungry.

My point is, the math is right but the answer it gives is a 2 megapixel image that could never be used to print from. Sorry, but I'm buying the D800 to print BIG, 30x48" specifically that one client requests.
So again, science right, but it took it to a level of an unusable photo.

--
-Dan
Winnipeg, Manitoba, Canada
'Cameras don't take pictures, people do.'
'No one sees your camera when they're looking at your pictures.'
http://www.danharperphotography.com/ -BLOG/stock site
http://www.danharperphoto.com/ -Commercial portfolio
http://www.wpgphoto.com/ -My Winnipeg based photography community

 

[Joofa]

Horshack wrote:

'* References:
http://forums.dpreview.com/forums/read.asp?forum=1018&message=39856801

I'm putting your references first as this topic comes up here so often. I think crames' post linked above is very good, as usual.

Theorem #1 - The 36MP sensor will have higher per-pixel noise than the 12MP sensor

True. (with the usual understood conditions.)

but both will have the equal amounts of noise on a per-area basis.

No, not necessarily. See, in digital we don't have areas, as in analog photography. We just have numbers. If we want to go from 36 million numbers to 12 million numbers, then some methods will yield poorer results, some the same, and others better. They will have equal amounts of noise if you downsample by binning. But, hardly anybody downsamples by binning in reality.

Consider the x2 downsampling in the following image to the right. Do they both seem to have the same SNR?

Theorem #2 - Downsampling does not increase the SNR of the image, at least across all spatial frequencies within the image.

Firstly, for small downsampling ratios the SNR will increase for natural images. Secondly, the way I see it, this notion of "all spatial frequencies" is for pure academic purposes only. Come to think of it. In your 16 MP pixel images, do you want to compare 16 million noise figures for each spatial frequency in the image! No body does that. As an example, when we say that read noise is say 2 e-, that means an avarage figure over all pixels in an image. Similarly, it is just bogus to have a list of millions of numbers in the frequency domain. We want just a few numbers, the less better, even in frequency domain to describe the overall noise content.

--
Dj Joofa

 

[SUPERHOKIE]

Good Lord, can you all just stop with this nonsense???

D800 down sampled to 12MP will have LESS noise and MORE details than the D700 does in its native 12MP resolution.

It's good to have some technical discussions once in a while, after all we are all die hard fanatics with photography, that's why we spend so much money on cameras and lenses, because we just love it!

But this is getting ridiculous now. All these charts and graphs, signal to nose ratio algorithms, noise derivative from down sampling at the 60HZ which is half the size of the pixel pitch of the light wave that can be captured by each pixel which has a diameter of 2.5 nanometer etc blah blah...

C'mon guys, I only have a D700 and will not be getting the D800 anytime soon because I don't need 36MP and I'm happy with the noise of D700 until ISO 6400. Frankly, I don't ever go above 6400. But it is pretty much a given that 36MP downsampled to 12MP will have better noise and details than a native 12MP image, from the samples we've seen so far.

Who really cares about minute technical details that does not matter that much?

Some of you are seeing the tree but missing the forest.

 

[ejmartin]

I'm not ee either (mathematics graduate). Can't this phenomenon be simply thought of a simple application of the central limit theorem?

Yes, together with the usual rule for the aggregation of variances (variance of the sum is the sum of the variances).

--
emil
--



http://theory.uchicago.edu/~ejm/pix/20d/

 

[ejmartin]

So your answer could explain why I get very good results if I use NR at full res ?

The NR also kills the high frequency noise but still leaves plenty of hi frequency detail in the picture, and more detail actually in the full res picture than in the one after resampling. In other words, could we say that high freq NR may be 'more efficient' at higher MP's?

Yes indeed. A common method of NR uses the 'wavelet decomposition', roughly a decomposition of the image into detail subbands of various scales (Nyquist to half Nyquist, half to one quarter Nyquist, etc). One reduces the magnitude of the wavelet coefficients that are smaller than some threshold under the theory that large coefficients represent true microcontrast while small ones are predominantly noise. Thus one retains more detail than would be obtained by downsampling, which would be equivalent to setting all the high frequency subband coefficients to zero (for a downsample by a factor two).

--
emil
--



http://theory.uchicago.edu/~ejm/pix/20d/

 

[David314]

So again, science right, but it took it to a level of an unusable photo.

let me try, same image, but opened the NEF in capture NX 2, turned noise reduction off, turned moire reduction on - neutral picture control

one at 36mpix one at 16mpix - cropped to same size

 

[David314]

one at 36mpix one at 16mpix - cropped to same size

sorry, i think i goofed and made it 12mpix downsized. at least i didn't have to make it 2mpix......

 

[Horshack]

Some relevant experiments:
Noise spectrum before/after downsampling:
http://forums.dpreview.com/forums/read.asp?forum=1000&message=30169513

Note in the power spectra, 256 on the horizontal axis is always Nyquist, so doesn't mean the same thing for the full size and reduced images -- the top half the power spectrum of the original is chopped away under downsampling and is simply absent in the downsampled image, because Nyquist is half in terms of lines per picture height.

Reading the responses on this thread and then re-reading your link above finally drove the point home for me. If I had gone directly to your noise tutorial at http://theory.uchicago.edu/...ejm/pix/20d/tests/noise/noise-p3.html#pixelsize it would have clicked a lot faster; you explain the concept there in so many different and compelling ways it becomes almost impossible not to comprehend. Your help and contribution are greatly appreciated.

 

[Horshack]

My estimate is that the situation will be a lot worse for the "high resolution" model at higher ISOs, because the RN is a much larger part of the image noise power here. You need VERY low RN figures to compete with at least reasonably sized pixels at very high ISOs. For the foreseeable future, images from lower MP cameras will be more "manageable" in post at the highest ISOs. With smart NR, the high-MP-cameras can match them on almost equal basis, but that includes a lot more processing power, or enough MP to be able to do real hardware binning.

Thanks for running this experiment. I agree about the color blotchiness. It's something I've observed in the D800 samples I've applied NR/downsampling to. It almost approaches mottling at times. Do you think this is primarily a function of the demosaicing algorithm and if so, can it be addressed without too many side effects? Are there any post-demosaic strategies to address it?

 

[RP McMurphy]

My estimate is that the situation will be a lot worse for the "high resolution" model at higher ISOs, because the RN is a much larger part of the image noise power here. You need VERY low RN figures to compete with at least reasonably sized pixels at very high ISOs. For the foreseeable future, images from lower MP cameras will be more "manageable" in post at the highest ISOs. With smart NR, the high-MP-cameras can match them on almost equal basis, but that includes a lot more processing power, or enough MP to be able to do real hardware binning.

Thanks for running this experiment. I agree about the color blotchiness. It's something I've observed in the D800 samples I've applied NR/downsampling to. It almost approaches mottling at times. Do you think this is primarily a function of the demosaicing algorithm and if so, can it be addressed without too many side effects? Are there any post-demosaic strategies to address it?

Personally if I have heavy colour noise I apply surface blur in PS to the A and B channels in LAB

The lightness channel, where all the detail is, is untouched but the 2 colour channels can take a hell of a lot of blurring before being noticeable

The lightness channel can have some masked sharpening and the final image looks completely natural without any plastic look but huge amounts of NR in effect

If it's moderate I just use the LR NR finction and only do this is the noise is very objectionable

For me this gets rid of the mottling look and loses no fine detail - noise is still present in lightness channel but the image in no way looks noisy

My intention (having finally come round to getting a D800 over 2nd hand D700) would be to apply this approach before downsampling

I figure that high frequency detail (and noise) can be reduced as part of NR (but no so much in method above) but the mottling wont move because it's too big to be removed by the NR and effectively becomes preserved detail (and possibly exaggerated by subsequent downsampling operations)

The other element of course is to expose to the right

Detail in shadow areas is heavy in noise because it's getting less signal information. Capturing more information by exposing to the right should mean that there's more detail in shadows to pull out with standard LR tools (of course the image may have high DR and at high ISO you lose lots of DR so ETTR isn't always possible without blowing highlights)

I figure that with improvements in the sensor and the multitude of tools available, high so noise can be very effectively dealt with now - downsampling for me will be used after initial colour NR (via the LAB channel if too objectionable)

 

[danharperphotography.com]

let me try, same image, but opened the NEF in capture NX 2, turned noise reduction off, turned moire reduction on - neutral picture control

one at 36mpix one at 16mpix - cropped to same size

BUT, but but but, you cheated! You turned on moire reduction and used a picture control. The principal here was to see if just downsizing the photo would help reduce it's noise levels.
We all know that software can drastically help improve noise reduction.

But the...actually...I'm done here. This is a great discussion but it just doesn't matter to me anymore. Whatever method you use to make your photos look the best they can should be up to you. I'm getting the D4 in the next couple of days and I'm just gonna take some pics and enjoy. And then the next week, I'll get the D800 and do the same.
The cure for the DPR blues is:
-Take photo
-Repeat.


--
-Dan
Winnipeg, Manitoba, Canada
'Cameras don't take pictures, people do.'
'No one sees your camera when they're looking at your pictures.'
http://www.danharperphotography.com/ -BLOG/stock site
http://www.danharperphoto.com/ -Commercial portfolio
http://www.wpgphoto.com/ -My Winnipeg based photography community

 

[lock]

If with high MP's the hi freq noise is best handled with good NR software (it works even after upsamlpling according to Dominique),
and downsampling does nothing to mid and low frequency noise,
than what's the gain in downsampling other than getting a smaller file size ?

 

[RP McMurphy]

In theory the downsaampling should aggregate out the random high frequency noise by downsampling but leave the detail - in that detail is likely to have neighbouring pixels replicating data and so the downsampling program will likely hold more of the repetitive high frequency detail and remove some high frequency random info

I guess there's more learning to do re high v mid frequency noise

Personally I think that downsampling is bound to help but appreciate that some downsampling programs may misread mid frequency 'noise' for detail and possibly worsen the noise in some cases

Most of this theory is beyond me but I'm just trying to pick up what I can that I may be able to apply to the D800 when it arrives

You dont need to downsample but of course you will not display at full size every time and therefore there will be some downsampling occurring somewhere at output stage

 

[wombat779]

Ok I see where you're coming from and you're totally right, reducing the image, reduces the visible noise, I'll give you that. But you had to reduce it to an unusable size to make it work to your point.

I don't think anyone has claimed that downsampling a particularly GOOD method of noise reduction, just that it is one method. The only useful purpose I can see of downsampling without other noise reduction is to facilitate comparisons to other lower resolution cameras at the pixel level (e.g. to compare a D800 image to a D700 or a D4 image), as downsampling will roughly simulate the impact on noise from the larger pixel size of the lower resolution cameras.

I think a much more effective method is to apply moderate noise reduction to an image before downsampling. While the noise reduction will destroy some detail, it is detail that would otherwise be lost in the downsampling process anyhow. Thus, you are basically using the relatively "smart" noise reduction in your PP software instead of the relatively "dumb" noise reduction produced by downsampling alone. Using this method, I can get a quite clean and detailed 12MP or 16MP image from a noisy 36MP image.

 

[David314]

let me try, same image, but opened the NEF in capture NX 2, turned noise reduction off, turned moire reduction on - neutral picture control

one at 36mpix one at 16mpix - cropped to same size

BUT, but but but, you cheated! You turned on moire reduction and used a picture control.

The ONLY difference between the two images is the size

They BOTH use the same picture control, NR off, and moire reduction

I did that because I thought your d800 fullsize image looked terrible

If you like, I can change all that back and you will still see a difference in the noise

The principal here was to see if just downsizing the photo would help reduce it's noise levels.

We all know that software can drastically help improve noise reduction.

But the...actually...I'm done here. This is a great discussion but it just doesn't matter to me anymore. Whatever method you use to make your photos look the best they can should be up to you. I'm getting the D4 in the next couple of days and I'm just gonna take some pics and enjoy. And then the next week, I'll get the D800 and do the same.
The cure for the DPR blues is:
-Take photo
-Repeat.

 

[ejmartin]

If with high MP's the hi freq noise is best handled with good NR software (it works even after upsamlpling according to Dominique),
and downsampling does nothing to mid and low frequency noise,
than what's the gain in downsampling other than getting a smaller file size ?

Not much. But those were the parameters of the discussion.

--
emil
--



http://theory.uchicago.edu/~ejm/pix/20d/