Pages: [1]   Go Down

Author Topic: image sensor noise test  (Read 1711 times)

torger

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 3267
image sensor noise test
« on: May 14, 2012, 03:56:31 am »

I've done a simple visual image sensor noise test that some of you may find interesting:

http://www.ludd.ltu.se/~torger/photography/noise-test.html

It compares Nikon D7000 (did not have a D800 unfortunately, but per pixel they are almost identical), a Canon 5D mark 2 (is its DR as bad people say?) and a Leaf Aptus 75 (has medium format as good DR as they say?). I'd love to have an IQ180 to test state-of-the-art MF too but I did not feel like renting one just for this simple test. The older Aptus is more relevant to those of us that are amateurs anyway, because these are sort of affordable on the second hand market.

I also compare the result in my test to the numbers found on DxOMark and discuss the differences.
Logged

hjulenissen

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 2051
Re: image sensor noise test
« Reply #1 on: May 14, 2012, 07:54:00 am »

Thank you for an interesting read.
Logged

BJL

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 6600
Re: image sensor noise test
« Reply #2 on: May 14, 2012, 09:13:46 am »

Thank you Anders,

    These tests seem to confirm and somewhat quantify several predictions about sensor performance comparisons, in particular the advantage in moderate shadows of a larger sensor due the fundamental advantage of being able to gather more photons and thus diminish the effect of shot noise, even while the smaller sensor with less dark noise has an advantage in deep enough shadows and when the sensors are severely "underexposed" (meaning, used at an exposure index well above their ISO base sensitivity).

One suggestion: for sensors like the MF CCDs and Sony/Nikon CMOS, the noise seems to enternthe signal almost entirely before any gain is applied, so that the DXO measures of SNR on an 18% gray card at various exposur index settings can be reinterpreted as measuring the SNR in darker parts of the image, at least as a rough guide. For example, the ā€¯midtone SNR" given there at exposure index 1600 is probably a good estimate of SNR in a shadow region three stops below the midtones at EI 200.

But beware: the DXO calibration of "ISO speed" is complete nonsense when applied to those MF CCDs, so it is better to look at the manufacturer's stated speed ratings. The problem is that when a DMF back takes advantage of its 16-bit ADC to position the midtones further down from maximum signal (while still having them at roughly comparable numerical levels, but with the maximum numerical level being four times higher than with 14-bit output), DXO absurdly calls this a two-stop over-statement of ISO sensitivity. This is due to DXO confounding a definition of base ISO speed relative to saturation of electron wells at photosites with "saturation" of the levels in ADC output. The bottom line is that the DXO curves for DMF backs should effectively be pushed about two stops to the right before making comparisons to smaller format sensors.

This DXO quirk has far smaller effects between other DSLRs, which generally do not have the same great surfeit of ADC levels relative to the dynamic range of the signal being digitized, and so tend to position the midtones at roughly the same fraction of maximum ADC output level rather than leaving an abundance of headroom.
« Last Edit: May 14, 2012, 01:30:37 pm by BJL »
Logged

bjanes

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 3387
Re: image sensor noise test
« Reply #3 on: May 14, 2012, 10:13:46 am »

I've done a simple visual image sensor noise test that some of you may find interesting:

http://www.ludd.ltu.se/~torger/photography/noise-test.html

It compares Nikon D7000 (did not have a D800 unfortunately, but per pixel they are almost identical), a Canon 5D mark 2 (is its DR as bad people say?) and a Leaf Aptus 75 (has medium format as good DR as they say?). I'd love to have an IQ180 to test state-of-the-art MF too but I did not feel like renting one just for this simple test. The older Aptus is more relevant to those of us that are amateurs anyway, because these are sort of affordable on the second hand market.

I also compare the result in my test to the numbers found on DxOMark and discuss the differences.

Anders,

I would like to compliment you for a very well done post, but take exception with your statement, "To sum up I would say that DxOMark does not provide any really good measurement for practical dynamic range, the dynamic range diagram digs too deep into the noise, and the SNR 18% is too bright light level to show any significant differences". One can use the full SNR plot (shown below for the D7000) and calculate the DR for any given noise floor as explained by Emil Martinec in this post.

I calculated the DR for each camera at a noise floor of 18 dB (approximately 8:1) using this method. For the D7000, the gray scale value for this noise floor is 0.18%, corresponding to a DR of 548:1 or 9.10 stops. The corresponding values for the Leaf and Canon are 8.14 and 7.99 stops respectively. One can normalize these values for resolution (print DR in DXO terminology) using the DXO formula. I get normalized values of 9.62, 9.17 and 9.02 stops for the Nikon, Leaf, and Canon respectively. These values are in general agreement with your observations.

Regards,

Bill
« Last Edit: May 14, 2012, 10:16:41 am by bjanes »
Logged

torger

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 3267
Re: image sensor noise test
« Reply #4 on: May 14, 2012, 10:24:50 am »

Anders,

I would like to compliment you for a very well done post, but take exception with your statement, "To sum up I would say that DxOMark does not provide any really good measurement for practical dynamic range, the dynamic range diagram digs too deep into the noise, and the SNR 18% is too bright light level to show any significant differences". One can use the full SNR plot (shown below for the D7000) and calculate the DR for any given noise floor as explained by Emil Martinec in this post.

I calculated the DR for each camera at a noise floor of 18 dB (approximately 8:1) using this method. For the D7000, the gray scale value for this noise floor is 0.18%, corresponding to a DR of 548:1 or 9.10 stops. The corresponding values for the Leaf and Canon are 8.14 and 7.99 stops respectively. One can normalize these values for resolution (print DR in DXO terminology) using the DXO formula. I get normalized values of 9.62, 9.17 and 9.02 stops for the Nikon, Leaf, and Canon respectively. These values are in general agreement with your observations.

Thanks for the feedback. I'll update the text with tips about the full-range diagram (which unfortunately cannot be brought up in the comparison mode?). I still think it is a bit limited though, the 5D mark 2 loses subjectively quite some range due to the patterned noise issue which cannot be seen in an SNR diagram.
Logged

bjanes

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 3387
Re: image sensor noise test
« Reply #5 on: May 14, 2012, 11:38:54 am »

Thanks for the feedback. I'll update the text with tips about the full-range diagram (which unfortunately cannot be brought up in the comparison mode?). I still think it is a bit limited though, the 5D mark 2 loses subjectively quite some range due to the patterned noise issue which cannot be seen in an SNR diagram.

Quite true. Also, the pattern noise is most prominent in the deep shadows and its significance will depend to some extent on the image content--i.e how many deep shadows are in the image. Furthermore, the noise floor for photographic DR is subjective. Everyone agrees that 1:1 is too low. 8:1 may be acceptable to some, while others may want a higher value. What is acceptable is best determined by visual inspection of the rendered image and this is why visual tests such as yours are complementary to the statistical approach of DXO.

Regards,

Bill
Logged
Pages: [1]   Go Up