Saturday, August 24, 2013

Bottle of French Wine for the Explanation of Data

Albert Theuwissen offers a bottle of good French wine to someone who explains the lack of consistency between gain and exposure settings in his CMOS sensor measurements. One can expect that signals measured at gain=1, exposure=42.24ms and gain=4, exposure=10.56ms are equal. In fact, they are not:

Sensor output value (corrected for the offset) as a function of
measurement number, for each measurement the camera gain
and exposure time are matched to each other.
Same data as above, but now with the regression lines added.

16 comments:

  1. Dark current? There's more time for dark current at longer integration and it is also non-linear with time.

    E

    ReplyDelete
  2. I think dark current has already been taken care of when he mentioned offset being corrected. IMHO this is caused by nonlinearity of some kind, which could be in the photodiode, the FD, or the gain. Albert, please repeat the experiment with fixed gain (but different exp), and fixed exp (but different gain), then see which one is non-linear.

    ReplyDelete
  3. I'd guess that the exposures are not accurate. Probably the actual exposure goes to zero before the exposure setting goes to zero. There isn't enough data to separate this effect from ADC offset but one more point taken in the dark will solve that.

    ReplyDelete
  4. If the exposure is set by a pulsed light source, it is possible that the longer pulses have larger integrated flux due to on/off risetimes or higher brightness with longer pulses or even a small spectral shift due to thermal effects in the source. The source needs to be observed with a photodiode with a sub-millisecond risetime to determine the relative total energy in short vs. long pulses.

    ReplyDelete
  5. Here is an experiment. Use a pulsed LED source and set it for roughly 50% duty cycle with a 10 ms period. Let it run continuously. Set the sensor exposure to capture 2 or 8 pulses to approximate your original test. The camera should be synced to the light pulse train so that the exposure starts and ends during the light off periods. See if you get the same problem. If so, I hereby turn the problem over to the sensor internals experts.

    ReplyDelete
  6. The answer is simple but giving it depends on the wine quality

    ReplyDelete
  7. It's the finite rise time of the led source

    ReplyDelete
  8. some of processing chain offsets might be affected by (analog?) gain...

    ReplyDelete
  9. Perhaps the inconsistency is due to oversampling of the French wine prior to doing the experiments.

    Seriously, I think it could easily be residual slightly negative offset getting gained up, or inexact gain, or a combination. The time dependence could be many many things.

    It cannot be that important if only French wine is being offered. Now if it was Californian wine, that would be a different matter.

    ReplyDelete
    Replies
    1. Hope that French wine not be the interpolation of the wine tasted prior to the experiments

      Delete
  10. Californian wine? What's that...

    ReplyDelete
  11. This comment has been removed by the author.

    ReplyDelete
  12. Albert,
    Two possible explanation (assuming switched cap amplifiers) (maybe both simultaneously)

    1) It could be the settling time of gain amplifier is affected due to higher gain leading to less settled output, while this may not necessarily affect noise,( if jitter is well managed) it certainly cause "less" output at higher gain.

    2)At unity gain, the capacitance in the feedback is the highest,--> best matching. Not so the case at gain = 4 when capacitance is 4 times smaller and more easily affected by bottom plate parasitics and mismatches.

    Hope we can toast.

    ReplyDelete
  13. It is a stupide question. How can we answer it without any details! Certains try to "answer" it as it were a serious question!

    ReplyDelete
    Replies
    1. Well, the details are in his blog. You are too lazy to read them though...have you been drinking too much of cali-forniated wine lately??!

      Delete
  14. Dark current is significantly less as expressed as a percentage of photons of interest at the shorter exposure. Unless the background read was done for same exposure period via shuttered event, dark current would contribute to final offset.

    The wine issue is entirely wrong, A nice white, or even better a Riesling would come from the Finger Lakes of NY :)

    ReplyDelete

All comments are moderated to avoid spam.