[Elphel-support] the discussion on the Heptaclops camera

support-list support-list at support.elphel.com
Tue Oct 30 10:47:53 PDT 2012


Hi Biel, have you read the article - http://www3.elphel.com/linuxdevices/articles/AT9913651997.html? And maybe also about the shot noise - http://en.wikipedia.org/wiki/Shot_noise ?

---- On Tue, 30 Oct 2012 09:25:24 -0700 Biel Bestué de Luna  wrote ---- 

>Andrey, about only having 150 gray levels, why does the camera only output 150 gray levels when the image that comes out of the sensor has 4096 output values? maybe this is because this amount of sampled information contain a lot of noise?

You know, when they have polls to measure public opinion on something - they have "margin of error"?  The totla group of electrons "polled" in a pixel is limited to the "full well capacity" - in our case ~8500. ADC has just a little less precision - can count by groups of two persons (4096 levels). So even if you had ideal 13-bit ADC (and zero noise preamplifier) that could count each single person (electron) and you can get measurement where "each vote counts", the difference between 8000 and 8001 is meaningful for democracy, but not a statistically meaningful difference if you try to measure public opinion. So depending on the "margin of error" you are willing to accept, there could be certain number of polling results that are different (by at least that "margin of error) from each other, and theer will be much less than 8500 or even 4096 outcomes from such public opinion measurements. And yes, you can use just 8 bits to represent such results.

>and, if 8 bits allow 256 shades of gray, does that meant that, while 8 bits per pixel allows it the camera can't reach the 256 shades of gray? if so, wouldn't it mean having a lot of banding?

When the quantisation step is safely smaller than the noise, there will be no banding because even the sensor looks at perfectly evenly illuminated surface, you will not get the same output code, because pixel-to-pixel variations (from the same noise) will be larger than the output code levels


>what factors need to change in order to have a camera that at least has the full 256 colours from 8 bits from the sensor?

It is good to have more quantization level than the number of S+N levels you can discriminate. Quantization noise add to the total noise, just when it is smaller than the sesnor noise, it's contribution is smaller in the total.

> maybe a sensor with higher ADC would do the job?

ADC is trying to measure number of electrons, so there is a natural limit to the ADC resolution needed.

>about shot noise, what factors make shot noise optimal for gamma 0.5, can this factors be changed for other gamma values? like 0.46 (1/2.2)?

Why do you need 0.46? Shot noise (predominant source of noise when you have good ADC and analog circuitry) is Math.sqrt() or Math.pow(x, 0.5). It is about preserving sensor information. When you want different contrast on the result video - that would be a separate story - you can apply it in post-processing, in the video editor - not in the camera.

Andrey
 






More information about the Support-list mailing list