Sharp's new sensor - just curious

[D Q E]

35mm film was used for cinema projection...

Was?

--
mumbo jumbo

 

[paulbod3]

Lost in a far future !

 

[alanr0]

I think then that even more jots per pixel are required. For
example, at 1/10th micron jot pitch, 6400 jots fit in the space of
a 8 um x 8 um pixel. Even denser jot pitches would be desirable.

So, what 12.5ish bits? What's the advantage over a conventional
construction?

The response is non-linear above 50% exposure (3200 of 6400 jots). I suspect dynamic range would be limited by second order effect in the silicon rather than maximum count. Perhaps 20 stops dynamic range? I would need to think harder about the noise statistics to come up with an upper limit.

Cheers.
--
Alan Robinson

 

[paulbod3]

Sorry, sometimes I am on my Digital Starship Zuiko in year 2525 and forget what is the present technology level available in cinemas.

35mm is still in use, the new standards have already been set but not yet of general use.

 

[Roland Karlsson]

What's the advantage over a conventional
construction?

1. You sample at very high spatial frequency so you don't have to worry about aliasing as the lens is the AA filter. This is true even for CFA (e.g. Bayer) type of color sensors.

2. You get extremely low noise at low light levels (if it works that is .

3. You get compression of the high lights leading to very high dynamic range.

4. You get an extremely simple logic, e.g. the shift out registers can be purely digital, i.e. adding no noise.

Personally I like this sensor VERY much (if it works that is .

The only problem (except maybe not working I can see is that the sensor has a fixed ISO sensitivity, only depending on the jot size. If you over expose you get compression of highlights and if you under expose you lose dynamic range. This "problem" could be fixed with the "grain" construct, increasing the ISO. But ... I am not sure this is a good idea as you lose both resolution and adds noise.

--
Roland

 

[paulbod3]

35mm film might have another 10 or more years life, considering the investment needed to substitute the machinery, and that we are now at 2k, 4k, probably 16k is the level needed.

For 16k the sensors would need a very tight pitch, so this discussion is quite valid in that field too, the change will need time but 2k and 4k are ready.

 

[D Q E]

35mm film might have another 10 or more years life, considering the
investment needed to substitute the machinery, and that we are now
at 2k, 4k, probably 16k is the level needed.

4K is already more than enough. Frankly, I feel 2K is plenty if the post chain is optimised.

For 16k the sensors would need a very tight pitch, so this
discussion is quite valid in that field too, the change will need
time but 2k and 4k are ready.

Ready? I work in the industry and 4K is hardly 'ready', it's a seriously niche workflow at the moment.

--
mumbo jumbo

 

[D Q E]

What's the advantage over a conventional
construction?

1. You sample at very high spatial frequency so you don't have to
worry about aliasing as the lens is the AA filter. This is true
even for CFA (e.g. Bayer) type of color sensors.

This is possible with existing designs.

2. You get extremely low noise at low light levels (if it works
that is .

Would you not also get very poor tonal range?

3. You get compression of the high lights leading to very high
dynamic range.

Again, both possible and not necessarily desirable at the moment

4. You get an extremely simple logic, e.g. the shift out registers
can be purely digital, i.e. adding no noise.

If you read every jot, wouldn't the clock have to be at a VERY high frequency?

--
mumbo jumbo

 

[D Q E]

I think then that even more jots per pixel are required. For
example, at 1/10th micron jot pitch, 6400 jots fit in the space of
a 8 um x 8 um pixel. Even denser jot pitches would be desirable.

So, what 12.5ish bits? What's the advantage over a conventional
construction?

The response is non-linear above 50% exposure (3200 of 6400 jots).
I suspect dynamic range would be limited by second order effect in
the silicon rather than maximum count. Perhaps 20 stops dynamic
range? I would need to think harder about the noise statistics to
come up with an upper limit.

Wouldn't the response curve be ALL knee, in fact? Why would you get a knee starting at 50% exposure? Doesn't every exposed jot decrease the remaining exposable area?

--
mumbo jumbo

 

[alanr0]

I think then that even more jots per pixel are required. For
example, at 1/10th micron jot pitch, 6400 jots fit in the space of
a 8 um x 8 um pixel. Even denser jot pitches would be desirable.

So, what 12.5ish bits? What's the advantage over a conventional
construction?

The response is non-linear above 50% exposure (3200 of 6400 jots).

Wouldn't the response curve be ALL knee, in fact? Why would you get
a knee starting at 50% exposure? Doesn't every exposed jot decrease
the remaining exposable area?

Yes, its non-linear all the way up, but the non-linearity becomes a lot stronger above 50%.

After thinking about this, I don't believe the dynamic range is as good as I first suggested, but it still shows an interesting 'soft clipping' characteristic. The fraction of unexposed jots decays exponentially, so:
Fraction of jots blackened (after exposure H) = 1 - exp(-H)

This saturates more quickly than I originally guessed, but still gives 3 stops headroom above the '50% exposed' level.

Cheers.
--
Alan Robinson

 

[Joe0Bloggs]

Film grains need to be hit a number of times before they will expose. These jots seem to expose to a single photon--or it they don't, they are subject to a QE factor, which makes them as likely to expose to the first photon as the next. Thus even if you have 100 small grains instead of one big grain in the same space, the likelihood that at least one grain will be exposed by a given number of photons is the same in both cases, assuming equal fill factor.

Hence ISO is variable without penalty just as in conventional sensors, with the caveat that higher jot density may also mean lower fill factor.

...my 2 cents

 

[paulbod3]

Up to what size of screen, in a normal theater, do you think the 2k is enough, and in what sense do you think the chain should be optimized ?

Actually the old Dof (zeiss) calculations were based on approx 1mp, that is in cinema if we think that a number of images is superimposed in our eye/mind could equate 3 mp in photo.

 

[Roland Karlsson]

1. You sample at very high spatial frequency so you don't have to
worry about aliasing as the lens is the AA filter. This is true
even for CFA (e.g. Bayer) type of color sensors.

This is possible with existing designs.

Not really - the readout noise will make the image very noisy.

2. You get extremely low noise at low light levels (if it works
that is .

Would you not also get very poor tonal range?

The tonal range would be by far superior to todays sensors.

3. You get compression of the high lights leading to very high
dynamic range.

Again, both possible and not necessarily desirable at the moment

Do you like hard clipped highlights? Do you like the noisy result when you have to under expose tho avoid clipped highlights?

4. You get an extremely simple logic, e.g. the shift out registers
can be purely digital, i.e. adding no noise.

If you read every jot, wouldn't the clock have to be at a VERY high
frequency?

Yes - but being pure digital shift registers you can clock very fast. You could also add one hardware counter for each pixel instead and just transporting the actual digital data.

--
Roland

 

[paulbod3]

Noise would be "localized", and processed at grain level, also offsetting the sensitivity/noise treshold of the grains, the clipping again might be localized.

In a way, an evolution of the Fuji HR.

Also in digital, if we set at black the noise and white the clipped parts, we can mask the effects, and the redundant spatial information can be used to interpolate the color again.

 

[paulbod3]

Diffraction limit is probably an overrated mith, and present calculations proposed are in my opinion a bit too conservative, aldo considering that diffraction doesn't happen all over the sensor at the same time in the same way; so that an image might appear clean well over the theoretical limit.

Astronomers simply blend an high number of images to obtain an higher resolution one of planets stars etc.

The DO lenses might behave in different way, as well the "glass", the number of components, the sealant, the coating might affect the final performance a lot.

 

[paulbod3]

In digital, it might be an advantage the RGB collection, the different colors are sampled at different spatial locations (Bayer) or depht (Foveon). If the image is recomposed for the best contrast, this slight spatial offset might help to reduce the problem a lot.

It is like decomposing the light in a rainbow, and then recomposing the image, by shifting the different colors of the needed amount.

Actually, the digital film might allow to reconstruct the correct image considering the different wavelenght of colors using the redundant resolution.

 

[paulbod3]

At 0.1µ / jot, a shift of R + G + B images can be controlled with enough precision to reduce the problem a lot. Wavelenght of colors are beetween 350 and 750nm (0.3-0.7 µ), so it would be possible to consider at least four/five steps of correction (x3 colors), if not many more. All depends on the resolution needed.

 

[Roland Karlsson]

Hi,

I made a small simulation where I compared Eric's proposal and an ordinary sensor. Those are my assumptions:

Ordinary
-----------
Saturation: 32000 electrons
Readout noise: 20 electrons
Photon noise: sqrt(electrons)
Base ISO: 100

Jot solution
----------------
Jots per pixel: 32000
Readout noise: none - this is the crucial, possible false, assumption
Photon noise: sqrt(electrons)

I assume the same QE for both (this assumption might be challenged). I assume that the lower limit is when the signal equals the noise in strength. DR is expressed in stops. The lack of readout noise increases the DR with 3 stops, i.e. you can use 2 stops higher ISO. The non linearity will increase the use of low ISO with 4 stops.

This is the result (sorry for the unreadable table):

-------------------------------------------------
ISO : DR/normal : DR/jot
-------------------------------------------------
12 : - : 15
25 : - : 14
50 : - : 13
100 : 9 : 12
200 : 8 : 11
400 : 7 : 10
800 : 6 : 9
1600 : 5 : 8
3200 : - : 7
6400 : - : 6
12800 : - : 5
-------------------------------------------------

So - instead of an ISO from 100-1600 we now have 12-12800. And instead of a maximum of 9 stops DR we now have a maximum of 15 stops DR. I think this is pretty impressive.

--
Roland

 

[Roland Karlsson]

OK -- 10 seconds after posting I see that I wrote several typos. OK - we do it again!!!!!

Hi,

I made a small simulation where I compared Eric's proposal and an ordinary sensor. Those are my assumptions:
Ordinary
-----------
Saturation: 32000 electrons
Readout noise: 20 electrons
Photon noise: sqrt(electrons)
Base ISO: 100

Jot solution
----------------
Jots per pixel: 32000
Readout noise: none - this is the crucial, possible false, assumption
Photon noise: sqrt(electrons)

I assume the same QE for both (this assumption might be challenged). I assume that the lower limit is when the signal equals the noise in strength. DR is expressed in stops. The lack of readout noise increases the DR with 3 stops, i.e. you can use 3 stops higher ISO. The non linearity will increase the use of low ISO with 4 stops.
This is the result (sorry for the unreadable table):

-------------------------------------------------
ISO : DR/normal : DR/jot
-------------------------------------------------
6 : - : 16
12 : - : 15
25 : - : 14
50 : - : 13
100 : 9 : 12
200 : 8 : 11
400 : 7 : 10
800 : 6 : 9
1600 : 5 : 8
3200 : - : 7
6400 : - : 6
12800 : - : 5
-------------------------------------------------

So - instead of an ISO from 100-1600 we now have 6-12800. And instead of a maximum of 9 stops DR we now have a maximum of 16 stops DR. I think this is pretty impressive.

--
Roland

 

[Roland Karlsson]

NOTE: The table assumes you expose after the brightest part in the image and that all the dynamic range then is put in the dark parts of the image. If you need more dynamic range in the brighter parts, e.g. bright highlights, then you have to adjust the ISO numbers - but the DR figures will still be the same.

--
Roland