Ad
Ad
Ad
Pages: « 1 ... 3 4 [5]   Bottom of Page
Print
Author Topic: 16 bit dslr  (Read 8497 times)
hjulenissen
Sr. Member
****
Offline Offline

Posts: 1679


« Reply #80 on: February 02, 2014, 12:19:53 AM »
ReplyReply

Hi,

Audio CD-s recorded at 16 bits. But photography is based on light and light has a noise of it's own. So 16 bit is now 12-14 bit of signal and 2-4 bits of noise.
I am sure that there are granularity limits for sound waves as well. But they tend to be irrelevant because all known acoustic environments have so much noise (air condition, outside traffic,...).

Audio offers 24 bit AD and DA today, but I believe that the Effective Number Of Bits is 20 or so. Peer-reviewed blind listening tests have so far been unable to distinguish 16-bit from higher resolution formats.

-k
Logged
hjulenissen
Sr. Member
****
Offline Offline

Posts: 1679


« Reply #81 on: February 02, 2014, 12:27:33 AM »
ReplyReply

Meanwhile, worries about Photoshop's 15-bit limit are encroaching.  I'm sure way back when they never thought we'd have real 16-bit image data, and that they could steal a bit for some reason.  But it looks like the day of reckoning may have come.
I really don't get this. Todays computers have really fast floating-point units. x86 allows for SIMD operations that can take chunks of 4 (or Cool float values simulatneously and do stuff like multiply-add.

While the throughput of 32-bit float add is likely 1/2 that of 16-bit fixed-point add, fixed-point operations tends to introduce more operations, and certainly slow down the developement/testing. Just as importantly, whenever Intel introduce new SIMD goodies (such as mmx->SSE->AVX) they will typically do floating-point first, then introduce fixed-point one generation later.

-h
Logged
ErikKaffehr
Sr. Member
****
Offline Offline

Posts: 7427


WWW
« Reply #82 on: February 02, 2014, 02:08:22 AM »
ReplyReply

Hi,

Well, we see new raw converters using floating point, HDR has some kind of floating point, too. Some people may object to floating point lacking precision, the 1.99999 != 2.00000 syndrome.

Most image processing software is pretty old and probably has a lot of quite obscure code based on binary operations. The best way is probably to start from scratch.

Best regards
Erik



I really don't get this. Todays computers have really fast floating-point units. x86 allows for SIMD operations that can take chunks of 4 (or Cool float values simulatneously and do stuff like multiply-add.

While the throughput of 32-bit float add is likely 1/2 that of 16-bit fixed-point add, fixed-point operations tends to introduce more operations, and certainly slow down the developement/testing. Just as importantly, whenever Intel introduce new SIMD goodies (such as mmx->SSE->AVX) they will typically do floating-point first, then introduce fixed-point one generation later.

-h
Logged

BJL
Sr. Member
****
Offline Offline

Posts: 5130


« Reply #83 on: February 02, 2014, 12:04:59 PM »
ReplyReply

My guess is that the long-time trend towards having more, smaller photo-sites on sensors of a given size will continue as long as the floor noise levels per photo-site continue to reduce, so that per-pixels SNR will never get beyond about 16,000:1 (2^14), meaning that 14 bits will always be enough. In fact, with full well capacities apparently steady at about 1600 electrons per square micron (according to Roger Clark at http://www.clarkvision.com/articles/does.pixel.size.matter/#Unity_Gain) once pixel sizes get down to about 3.3 microns (bigger than the Sony RX100's 2.4 microns; just a bit smaller than the 16MP 4/3" sensors), the full well capacity will be about 16,000 e- or less, so an ideal 14 bit ADC could count the electrons exactly.

This seems an easier path than having fewer, bigger photo-sites that need 16-bit ADCs, because the electron counting is done with more parallelism: with the same exposure over the same total sensor area, more smaller photo-sites leads to counting the same total number of electrons but in more, smaller bundles, using more column parallel ADCs so that 14-bit rather than 16-bits is enough, which allows faster ADC operation.
« Last Edit: February 02, 2014, 12:53:51 PM by BJL » Logged
Ajoy Roy
Jr. Member
**
Offline Offline

Posts: 82


« Reply #84 on: February 03, 2014, 08:47:48 AM »
ReplyReply

We are assuming that as the sensel sizes go down the full well capacity will decrease. That need not be so. If we have 3D sensels charge accumulation layer is relatively deep, they can accommodate a lot more electrons. If the full well capacity increases to a million electrons then we may have 16+ bit data.
Logged

Ajoy Roy, image processing
ErikKaffehr
Sr. Member
****
Offline Offline

Posts: 7427


WWW
« Reply #85 on: February 03, 2014, 11:29:29 AM »
ReplyReply

Hi,

We don't need a million, 25000 would be fine ;-)

Best regards
Erik

We are assuming that as the sensel sizes go down the full well capacity will decrease. That need not be so. If we have 3D sensels charge accumulation layer is relatively deep, they can accommodate a lot more electrons. If the full well capacity increases to a million electrons then we may have 16+ bit data.
Logged

BJL
Sr. Member
****
Offline Offline

Posts: 5130


« Reply #86 on: February 03, 2014, 11:35:23 AM »
ReplyReply

We are assuming that as the sensel sizes go down the full well capacity will decrease.
I am not simply assuming with no basis: I cited evidence a a clear pattern for some years of roughly constant electrons per unit area of photo-site or rough constant well depth. Things might change, but it does seem quite likely that well depth will not increase much.
Logged
Ajoy Roy
Jr. Member
**
Offline Offline

Posts: 82


« Reply #87 on: February 04, 2014, 07:16:45 AM »
ReplyReply

Similar limitations in normal CMOS IC manufacturing led to vertical (or 3-d) design, hence it may soon percolate to the photo sensor design.
Logged

Ajoy Roy, image processing
Pages: « 1 ... 3 4 [5]   Top of Page
Print
Jump to:  

Ad
Ad
Ad