[opendtv] Re: Precision


Dan: "First of all, bit depth: 8 bit, 10 bit, 12 bit, higher? What is the
graduation difference the eye can see at a particular wavelength?"

Mark Schubin: "So your question should be how much SNR we want, not how
many gradations we need."

Dan:

Good point.  I appreciate your discussion so that I can better communicate
in this realm.  I didn't really feel good about using the term "graduation"
when I used it.  I used it based on a maximum level and a minimum level,
and based on the bit depth, there are more or less "graduations."  You
would think I studied chemistry.

I came up with a different question:  What is the maximum contrast ratio
the eye can see and what do we need to do to quantify it?  Perhaps
"contrast ratio" is not the proper term (is "dynamic range" the proper
term?)  What I mean is that there is a maximum light level the eye can
perceive (and still resolve) and a minimum level the eye can take at any
given iris/pupil.  Of course, the eye's pupil is dynamic to make the
dynamic range very great indeed.  If we take the absolute lowest light
level the eye can see while the pupil is wide open and the absolute highest
the eye can see while the pupil near closed, we have the highest contrast
available to the human eye.  Naturally, the eye could never be there, so
perhaps this is not a scenario we would ever deal with.  Perhaps we only
need to quantify the range from the minimum to the maximum an eye can see
at an overall average scene light level.  In any case, we will have a
minimum light level and a maximum light level.  (I am only dealing with
luminance right now and not individual wavelengths, which adds another
layer of complexity.)

Imagers (i.e., CCDs, etc.) are trying to widen that dynamic range to match
the human eye.  It will be a while before the imager can get there, but
certainly there are improvements.  (And I am not talking about processing
to compress the levels, I am talking about raw photons hitting the imager
and turning that into electrons and then it into a discrete number.)  So
our lowest light level (0 photons) there is a low number (probably not 0
but relatively close) and then a maximum where the electrons start spilling
over, a high number.  So at the A/D, if we sample at different bit depths
(should I use the term "depths"?), we have more discrete steps with higher
bits or less with lower (again, I used the term "graduates" where I meant
"discrete steps").  This is where I am curious as to how many bits we need.
I believe that the best imagers are at 14 bits today with 22 bits in the
internal processing of the camera.  I suppose that as imagers improve--the
lower limits improves (SNR/noise floor of the pixel) and the maximum
becomes higher (more electrons before full)--that the bit depth will need
to continue to go up.  But what is the theoretical maximum before the best
eye can see a difference (perhaps at any given pupil setting)?

There is a second part to this question:  What bit depth do we need to
transmit to get better pictures.  I think the best demonstration that 8
bits is not enough is when there is a smooth gradient.  There are clear
steps where there shouldn't be.  Has anyone done research on how many we
need before we perceive a difference?  Should I be asking what the maximum
SNR one can perceive is?

I am open to hearing what terms I should be using within my discussion so I
that I can better communicate.

I just thought of another question:  I assumed the CCD is linear.  Is it?
And a CMOS imagers?

Dan Grimes

Other related posts: