-
Welcome to the new forums! Please read this first. For known issues we are working to resolve, click here.
Let's finally resolve the D800 downsampling / NR questions
- Thread starter Horshack
- Start date
-
- Tags
- nikon d700 nikon d800
bobn2
Forum Pro
Yes. We can look at an image in the spatial domain or the spatial frequency domain, which is given by the Fourier transform of the spatial domain. Then the difference between the lowest spatial frequency (set by the sensor size) and the highest spatial frequency (set by the pixel size) is the bandwidth.
The trouble is with those diagrams is the 'noise' being a separate bit down the bottom, the noise is spread throughout not just down the bottom. If you cut the bandwidth, you cut the noise.I'm one level too removed in the abstract to get my head around this. I made a diagram depicting what I think you're saying. Is this correct:
Yes, it is abstract. The frequency domain is not a very 'natural' way of looking at things, but happens to much more easily yield useful results. Think of the 'bandwidth' describing the range of details that can be recorded, from very coarse to very fine.
OK, another way of looking at it is that the noise is in the structure of the light, so when you see 'noise', what you are seeing is the fine detail of the light. Since 'noise' is just detail, if you see more detail, you see more noise, if you see less detail you see less noise.
--
Bob
KenEis
Senior Member
Theorem #1 - The 36MP sensor will have higher per-pixel noise than the 12MP sensor but both will have the equal amounts of noise on a per-area basis. In other words, if you combine multiple 36MP pixels into units roughly equal in size to 12MP pixels, the amount of noise in those combined pixels will be equivalent to the 12MP pixels. This seems pretty obvious and straightforward.
--
Noise is random. Lets assume Your statement that two sensors have the same noise per area is correct. The noise from pixels that are to be combined would be uncorrelated so the combining would reduce the noise. You postulate from an assumption that the pixels you are combining all show the same noise (chroma and or luminousity. and in the same spatial distribution.
Assuming that your "combining" is just averaging the resulting pixel would have less noise. If I remember my information theory the noise will fall of as the square root of the number of pixels to be combined. So if you combine 4 noisy pixels you will reduce the noise by a factor of 2.
Ken Eis
http://keneis.zenfolio.com
--
Noise is random. Lets assume Your statement that two sensors have the same noise per area is correct. The noise from pixels that are to be combined would be uncorrelated so the combining would reduce the noise. You postulate from an assumption that the pixels you are combining all show the same noise (chroma and or luminousity. and in the same spatial distribution.
Assuming that your "combining" is just averaging the resulting pixel would have less noise. If I remember my information theory the noise will fall of as the square root of the number of pixels to be combined. So if you combine 4 noisy pixels you will reduce the noise by a factor of 2.
Ken Eis
http://keneis.zenfolio.com
Luke Kaven
Veteran Member
Yes, exactly!
Luke Kaven
Veteran Member
There could most definitely be high frequency signal that you don't want to throw away.
in the case of the D800, you can aggregate your photons according to whether photons are plentiful or scarce. When photons are plentiful, the D800 delivers good pixel-level performance with low noise at 36MP. When photons are scarce, you will tend to aggregate the photons into bigger bins to improve the overall noise profile.
Also, in low light cases, you can use multiple exposures to supersample the image space, and average them together. The random components of noise will drop out at a predictable rate, while the non-random details will become clearer. Not that this is always practical. I'm just trying to show that there are some ways to mitigate high frequency noise without throwing away the corresponding signal.
Luke Kaven
Veteran Member
In general, Lanczos (Lanczos-windowed SinC method) is very clean. Comes with rawtherapee. But bicubic works.
Bicubic
Your post has bridged a big gap in my comprehension, I think. Let me reprhase in my own relatable terms and please tell me if I've got it. A 36MP sensor will sample higher frequencies than a 12MP sensor. The higher frequencies it samples will include both higher-frequency detail and higher-frequency noise. If we devise a theoretically-perfect low pass filter which has no side effects like aliasing, we can take the 36MP sample and produce data that is identical to the native 12MP sensor sample, in terms of both signal and noise. I drew a diagram to depict this, even though it probably doesn't need one:
Red G8R
Senior Member
Just answer it simply for me. If we made identical prints (say 13x19) of identical scenes from both sensors, 12 mp vs 36 mp, same lens, jpegs, no pp, which one would show more noise?
For me, this is what will happen in practice.
--
Peter
Ontario, Canada
For me, this is what will happen in practice.
--
Peter
Ontario, Canada
Luke Kaven
Veteran Member
I can add one more thing to that.
The total MTF between (a) a capture from a native 12MP sensor, and (b) a capture from a native 36MP sensor downsampled to 12MP, will be different.
In an important sense, the native 12MP capture does not exploit 100% of the capacity of the 12MP sample space at the highest frequencies. This might be attributed to at least two things: (i) a difference in the OLPF between the low res and high res sensors, and (ii) bayer demosaic artifacts.
See for example this:
http://diglloyd.com/blog/2009/20090109_1-NikonD3x.html
After downsampling, system (b) will show more high-frequency detail than system (a).
The total MTF between (a) a capture from a native 12MP sensor, and (b) a capture from a native 36MP sensor downsampled to 12MP, will be different.
In an important sense, the native 12MP capture does not exploit 100% of the capacity of the 12MP sample space at the highest frequencies. This might be attributed to at least two things: (i) a difference in the OLPF between the low res and high res sensors, and (ii) bayer demosaic artifacts.
See for example this:
http://diglloyd.com/blog/2009/20090109_1-NikonD3x.html
After downsampling, system (b) will show more high-frequency detail than system (a).
Luke Kaven
Veteran Member
With sensors of equivalent noise ratings and different pixel counts, the results -- in terms of noise alone -- should be more or less equivalent. However, the downsampled 36MP capture will retain more detail at the highest spatial frequencies. See for example, this:
http://diglloyd.com/blog/2009/20090109_1-NikonD3x.html
Understood Luke. I have lots of questions relating to how demosaicing will affect the detail and noise for the theoretical 36MP downsample but I'm holding off on that until I firm up the basics. I guess I should have titled the thread 'Educating Horshack', since I obviously was starting at a lower level of understanding than others. ;-) But I'm ramping up quick and hopefully we can touch on things that will illicit fresh thinking about this.
ejmartin
Veteran Member
Some relevant experiments:
Analysis of FT spectrum for Neat Image acting on different spatial frequencies:
http://forums.dpreview.com/forums/read.asp?forum=1018&message=31774115
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.
--
emil
--
http://theory.uchicago.edu/~ejm/pix/20d/
Analysis of FT spectrum for Neat Image acting on different spatial frequencies:
http://forums.dpreview.com/forums/read.asp?forum=1018&message=31774115
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.
--
emil
--
http://theory.uchicago.edu/~ejm/pix/20d/
ejmartin
Veteran Member
Or maybe even elicit it, if we want to stay on the right side of the law
--
emil
--
http://theory.uchicago.edu/~ejm/pix/20d/
Thanks Emil, those depictions were very helped (I also consulted http://qsimaging.com/ccd_noise_interpret_ffts.html to help my understanding of your FFTs).
I have a question regarding the following diagram, which was generated by 'crames' in response to a post of yours http://forums.dpreview.com/forums/read.asp?forum=1018&message=39859119 )
I follow how the SNR at the original nyquist 'n1' is at or below 1:1 and that a downsample moves nyquist down to 'n2', causing the low SNR data between 'n1' and 'n2' to be eliminated, which increases the average SNR of the remaining image data. My question is, how does one quantify this? I've read elsewhere that a typical sensor's MTF is around maybe 70% of nyquest. Whatever the actual MTF, since the SNR is a slope from the original nyquist to the downsampled nyquist, what kind of function can be applied to derive the average SNR gain for the downsampled image? Also, what are the qualitative/perceptual differences of noise reduced near nyquist vs noise+signal reduce well below nyquist?
;-) You physics guys aren't supposed to know how to spell well too.Or maybe even elicit it, if we want to stay on the right side of the law
Greetings. The problem with these analyses is that the concept of noise is ill defined. Much of the discussion about noise in these discussions are about sensor measurements via raw files which are fundamentally different than the visual defects seen in images that we see as noise. Demosaicing in the raw conversion process will spread a single pixel error and have an impact on both luminance and chroma. Demosaicing algorithms will differ among conversion software (not to mention camera manufacturers) and differ most on how edges are handled. Different downsizing algorithms also will have differential impact on noise.
Why there is such disagreement about the imact of downsizing is your mileage may vary depending on what collection of sensor, demosaicing, downsizing, etc. one is using not to mention the characteristics of the "noise" under question.
IMO, downsizing doesn't obviously reduce noise. I've certainly seen cases (in my own downsizing) where it has increased noise (color, in particular).
Good luck with coming up with a single satisfactory answer to this question.
Cheers,
-Yamo-
Luke Kaven
Veteran Member
It's an interesting area, isn't it? I'm not an EE or physicist, but have long experience in other areas of science and philosophy. In other words, I have lots of holes, but a good grasp of inquiry in the area. So it is really helpful for me when people like Emil Martinec come around and just impart a lot of clarity into an otherwise muddled subject area. I owe much of what I "know" to his contributions, with help from other contributors like Newman, Oelund, Claff, and Sheehy.
I started in on this because I really am a low-light junkie and equal part night-owl. I've really learned a lot about how to get photographs that work this way.
Keyboard shortcuts
- f
- Forum