25 MP sensor of Sony

Chris Malcolm <cam DeleteThis @holyrood.ed.ac.uk> wrote:
>David J Taylor <david-taylor DeleteThis @blueyonder.neither-this-bit.nor-this-bit.co.uk> wrote:
>> Chris Malcolm wrote:
>> []
>>> They can be encoded with any number of bits from 1 upwards. You don't
>>> lose dynamic range with fewer bits, you lose fineness of gradation
>>> steps. With one bit all you get is white and black. With four bits you
>>> get two intermediate shades of grey. With 12 bits you get 4000 shades
>>> of grey. I'm pretty sure I can't distinguish 4,000 different shades of
>>> grey, so I'm sure 12 bits is more than enough for the dynamic range of
>>> my eyes, other things being equal. There's no reason why a 10bit ADC
>>> couldn;t cover 10 times the dynamic range of a 15 bit ADC. However,
>>> it's often technologically convenient to use more bits than necessary
>>> for a variety of reasons, and those reasons change with the
>>> technology.
>
>> Remember the effects of gamma correction, though. 255 shades of
>> gamma-corrected data could have a linear dynamic range of more than
>> 4000:1. Yes, you are unlikely to distinguish at the high end (3998 vs.
>> 3999), but the gamma-correction seen in JPEGs does make the lower end
>> important.
>
>You're quite right, the non-linearity of gamma correction does
>increase the resolutions required. The important point remains that
>the number of bits in the encoding doesn't affect the dynamic range
>between darkest and lightest, just the number of steps in between.

"Recording" two steps might at first glance appear to be
able to record a large dynamic range, but it can't. Two
very small steps would have exactly the same data (i.e.,
the SNR is half the signal level). It's what can be
extracted from the digital data, not what goes into it,
that determines if it has a large dynamic range.

Simply put, 14 bits allows recording a greater dynamic
range that 12 bits, which allow more than 10 bits. Try
all you like, that can't be escaped. The only question
is whether the analog input data actually has enough
dynamic range to require more bits.

Keep in mind that the sensors are analog, and linear.
The input to the ADC is analog, and linear. Gamma
correction is accomplished with the digital data *after*
the data (whatever it's dynamic range is) is recorded.
And if you record it with only 8 bits it will *not* have
the dynamic range that is useful with an 8 bit gamma
corrected image (i.e., a JPEG).

--
Floyd L. Davidson <http://www.apaflo.com/floyd_davidson>
Ukpeagvik (Barrow, Alaska) floyd DeleteThis @apaflo.com

Chris Malcolm <cam DeleteThis @holyrood.ed.ac.uk> wrote:
>Floyd L. Davidson <floyd DeleteThis @apaflo.com> wrote:
>>>There's no reason why a 10bit ADC
>>>couldn;t cover 10 times the dynamic range of a 15 bit ADC.
>
>> Wrong. Signal to Noise Ratio is the reason.
>
>>>However,
>>>it's often technologically convenient to use more bits than necessary
>>>for a variety of reasons, and those reasons change with the
>>>technology.
>
>> So just show us a case where that is being used by a
>> DSLR today. (It is very common for digital applications
>> using voice frequencies, such as telephones and modems.
>> It is not common, yet, with DSLR's.)
>
>>>Given Sony's different implementation of the ADCs their 12 bits might
>>>in fact offer less noise and more dynamic range than the usual 14
>>>bits, just as X's 12MP camera might offer higher resolution images
>>>than Y's 13MP camera. It's not just the numbers.
>
>> That's nonsense.
>
>There's no need to go into all that technical detail. The concepts are
>quite simple.

Apparently even *more* technical detail is required,
because while the concepts are indeed simple, they are
being missed:

>Any dynamic range can be represented by any number of
>bits simply by adjusting the step size.

That is false. A 1-bit system will have a dynamic range
of 7.782 dB. Each additional bit will increase the
dynamic range by 6.02 dB at most.

That is assuming the optimum step size (anything larger
will decrease the dynamic range) and sampling point.

What you are leaving out of the equation is quantization
distortion. For *any* 1-bit system you must first
define the step sizes, which will generically be +s/2
and -s/2, where s is the voltage increment of the step.
Since the signal is now defined as ranging from +s to
-s, and the error signal, e, is defined as the RMS value
of energy spread over that range, which works out to
0.408, and can be converted to dB with

20 * log (0.408) = -7.82 dB

Each additional bit will allow 6.02 additional dB of
dynamic range.

You can look up the numbers in just about any text on
the theory and practice of digitization.

>Reducing bit length of
>representation only reduces dynamic range if step size is held
>constant. What step size is deemed an acceptable minimum depends on
>the associated technologies and various technological compromises. The

The dynamic range is a function of the step size, not
the number of bits as such. If you make the step size
larger in comparison to the overall signal range, two
things happen: less dynamic range and fewer bits are
required.

In the end, if you make one step smaller you
*necessarily* make another step larger, within a given
range and using a given number of bits (i.e., a given
number of steps). That is "gamma correction".

The effect is to lower the signal to noise ratio in one
part of the signal range to make it higher in another.
That extends the *useful* dynamic range by allowing
smaller steps sizes at lower signal levels (and thus a
better SNR) while increase step sizes and thus reducing
SNR at higher signal levels.

>fact that increases in bit length of representation have been
>associated with increases in dynamic range is an accident of history
>and common technological compromises.

It is hardly an accident of history. It is indeed one
of the very common compromises, trading unnecessary SNR
in one part of the signal range for desired SNR
improvement in another.

All you are suggesting is that after gamma correction
the dynamic range is not necessarily reduced. But that
only happens after the linear data from a sensor is run
through an ADC to produce "linear raw data". It's true
that with over sampling it is possible to then later
reduce the number of bits used to represent a given
dynamic range. It's a trade off though, and sacrifices
maximum SNR and complexity.

Of course the "raw" file is still the higher bit file...

>There are differences in the Sony technology. Therefore you can't
>simply compare the numbers unless you have shown that the differences
>don't matter.

As near as I can tell this has nothing at all to do with
sensor technology. Regardless, you can rest assured
that Sony will not be going back to 10 bit ADCs...

>Another poster has already suggested some reasons why
>they do matter.

To change the dynamic range available via a linear data
file???

--
Floyd L. Davidson <http://www.apaflo.com/floyd_davidson>
Ukpeagvik (Barrow, Alaska) floyd DeleteThis @apaflo.com