It's about time. I'm sick of the lack of dynamic range of existing CMOS sensors—burned out 'picket fence' highlights and noisy bottom ends (at least film has a nice s-shaped transfer curve).
Hopefully, if this tech actually amounts to anything then it'll force all other sensor makers to complete at a similar level.
If this tech comes to pass, then for the first time we'll have photographic processes that are approaching the dynamic range of the human eye. That can only be good for photography/imaging.
Like you, I remain sceptical because I've seen no evidence of a new tech breakthrough when it comes to image sensors.
The question one has to ask is how did Canon achieve such a huge step forward/large dynamic range when it's eluded others for years.
Image sensors vary from manufacturer to manufacturer but let's face it when all's equal—same given area/pixel number/size etc.—then the differences between them are small. We'd expect this to be fact because manufacturing techniques, chip manufacturing, photolithography etc., are well developed processes and the competition fierce, so they're all pretty much cutting edge.
My suspicion is that it's been achieved principally with computational/algorithmic processing. If so, then it's essentially all smoke and mirrors.
It's hardware level algorithmic, but it's not exactly "smoke" given my reading.
Dividing the full sensor into regions and essentially counting photons as they arrive in order to limit exposure time once there's sufficient captured range and contrast per grid region is a neat trick.
Ideally it's what a skilled photographer attempts per frame with conventional film - here it's gridded results with meta data across the full sensor allowing for better blending of the very dark to very light areas.
What would be interesting here would be to d/load a proper raw dump of all the clipped grids with their exposure timings and to them be able to mess with full image composition from the raw capture.
"...but it's not exactly "smoke" given my reading."
OK, I need to reread it carefully. Note, this matter has popped up again here: https://intels news.ycombinator.com/item?id=34590094. I've had a bit of rave there along similar lines. Also, others there have made some interesting comments.
As far as I'm concerned, hardware level algorithmic is essentially [but not fully] back-end processing (see my comment on that page). I've been wary of such solutions for years and they go right back to Intel's horrible Pentium bug. This so-called algorithmic solution was hardly an algorithm as such, it was nothing other than a badly implemented lookup table. Intel cheated and they weren't even smart enough to copy out the 'answers' correctly into the table. Similarly, any onboard microcode requires careful analysis before being given a tick. Sure, this is a different situation but the modus operandi isn't necessarily so.
That said, I must be fair so I'll reserve further judgement until I see the details.
Let's be clear about what I mean by 'improvements' both here and on the other HN web page (clearly, what I'm about to say deals with complex matters and necessity requires me to grossly oversimplify them).
Essentially, there are two physical ways improvements can happen: (a) concoct the better use of physics to achieve an improved quantum efficiency—use, say, a different semiconductor that has a 'better' photoelectric work function etc., and, (b) improve the engineering/mechanics (use better microlenses, improved pixel/cell construction, amalgamating cells under low light levels etc. etc.).
The final method, (c), relies on processing/algorithms which can both interact with the physical systems/hardware as well as 'clean up' images during post-processing.
When I mention physics to improve 'quantum efficiency' I'm referring to a system whose 'output' for any given number of quanta on the input has a higher output over a previous system. That is, more electrons are directly liberated to do work than before.
In my opinion, this is is big one—that is, it's the one in most need of real improvement now! It is also the most difficult to achieve, and it's always the factor that's most skimmed over by manufacturers in any of these announcements. Right, sensor manufacturers do not like talking about this 'limitation' as it's essentially a given quantity (by virtue of the physical properties of materials), and for standard manufacturing processing techniques that means using doped silicon.
Let's put this into true perspective: animals including humans can detect (respond to) single incoming photons, even the very best of these silicon-based sensors can do nothing like that (we're still orders of magnitude† off the mark):
My argument is straightforward: any improvements that do not address the fundamental sensitivity problem at quantum level can hardly be said to to be game-changers. That said, any improvement is to be welcomed, but at the same time we do not want manufacturers bullshitting us with 'smoke and mirrors' as they so often do.
_
† Except for photomultipliers, but then it's rather impractical to use a separate PM for every individual pixel!
I understand all that, I've been in remote sensing and signal processing for some time now (mid 1980s or there abouts).
This is NOT a big fundemental change - but it is an advance in hardware capabilities in the sense that small sections of the larger sensor array can now individually self regulate to prevent over saturation.
This effectively means that the entire array can deliver improved dynamic range over would could be delivered before as now entire sections that would otherwise return maxxed out counts (and thus lose any clarity of detail within those sensor regions) will now return a spread of values (and a note that they almost hit saturation in a third of the full exposure time).
This is a better design that returns a big picture with more "actionable infomation" - and it's almost "pre algorithm" raw information as it's likely that sensor data and sub region exposure time meta data would be available via a custom RAW dump.
I agree with you, perhaps more fully than you intend, that all hardware level design decisions make up the first (or perhaps 0'th) level of 'back-end processing'.
I also approve of the approach here, one large sensor array is now many small independent sensor regions.
Done properly it becomes and improvement, much as James Webb having many small individually tunable receptor regions was one of many improvements over prior space telescopes.
6 comments
[ 4.5 ms ] story [ 27.9 ms ] threadHopefully, if this tech actually amounts to anything then it'll force all other sensor makers to complete at a similar level.
If this tech comes to pass, then for the first time we'll have photographic processes that are approaching the dynamic range of the human eye. That can only be good for photography/imaging.
The question one has to ask is how did Canon achieve such a huge step forward/large dynamic range when it's eluded others for years.
Image sensors vary from manufacturer to manufacturer but let's face it when all's equal—same given area/pixel number/size etc.—then the differences between them are small. We'd expect this to be fact because manufacturing techniques, chip manufacturing, photolithography etc., are well developed processes and the competition fierce, so they're all pretty much cutting edge.
My suspicion is that it's been achieved principally with computational/algorithmic processing. If so, then it's essentially all smoke and mirrors.
Of course, I hope I'm wrong.
Dividing the full sensor into regions and essentially counting photons as they arrive in order to limit exposure time once there's sufficient captured range and contrast per grid region is a neat trick.
Ideally it's what a skilled photographer attempts per frame with conventional film - here it's gridded results with meta data across the full sensor allowing for better blending of the very dark to very light areas.
What would be interesting here would be to d/load a proper raw dump of all the clipped grids with their exposure timings and to them be able to mess with full image composition from the raw capture.
OK, I need to reread it carefully. Note, this matter has popped up again here: https://intels news.ycombinator.com/item?id=34590094. I've had a bit of rave there along similar lines. Also, others there have made some interesting comments.
As far as I'm concerned, hardware level algorithmic is essentially [but not fully] back-end processing (see my comment on that page). I've been wary of such solutions for years and they go right back to Intel's horrible Pentium bug. This so-called algorithmic solution was hardly an algorithm as such, it was nothing other than a badly implemented lookup table. Intel cheated and they weren't even smart enough to copy out the 'answers' correctly into the table. Similarly, any onboard microcode requires careful analysis before being given a tick. Sure, this is a different situation but the modus operandi isn't necessarily so.
That said, I must be fair so I'll reserve further judgement until I see the details.
Let's be clear about what I mean by 'improvements' both here and on the other HN web page (clearly, what I'm about to say deals with complex matters and necessity requires me to grossly oversimplify them).
Essentially, there are two physical ways improvements can happen: (a) concoct the better use of physics to achieve an improved quantum efficiency—use, say, a different semiconductor that has a 'better' photoelectric work function etc., and, (b) improve the engineering/mechanics (use better microlenses, improved pixel/cell construction, amalgamating cells under low light levels etc. etc.).
The final method, (c), relies on processing/algorithms which can both interact with the physical systems/hardware as well as 'clean up' images during post-processing.
When I mention physics to improve 'quantum efficiency' I'm referring to a system whose 'output' for any given number of quanta on the input has a higher output over a previous system. That is, more electrons are directly liberated to do work than before.
In my opinion, this is is big one—that is, it's the one in most need of real improvement now! It is also the most difficult to achieve, and it's always the factor that's most skimmed over by manufacturers in any of these announcements. Right, sensor manufacturers do not like talking about this 'limitation' as it's essentially a given quantity (by virtue of the physical properties of materials), and for standard manufacturing processing techniques that means using doped silicon.
Let's put this into true perspective: animals including humans can detect (respond to) single incoming photons, even the very best of these silicon-based sensors can do nothing like that (we're still orders of magnitude† off the mark):
https://www.snexplores.org/article/our-eyes-can-see-single-s...
https://www.scienceabc.com/humans/can-human-eye-see-detect-s...
My argument is straightforward: any improvements that do not address the fundamental sensitivity problem at quantum level can hardly be said to to be game-changers. That said, any improvement is to be welcomed, but at the same time we do not want manufacturers bullshitting us with 'smoke and mirrors' as they so often do.
_
† Except for photomultipliers, but then it's rather impractical to use a separate PM for every individual pixel!
;-)
This is NOT a big fundemental change - but it is an advance in hardware capabilities in the sense that small sections of the larger sensor array can now individually self regulate to prevent over saturation.
This effectively means that the entire array can deliver improved dynamic range over would could be delivered before as now entire sections that would otherwise return maxxed out counts (and thus lose any clarity of detail within those sensor regions) will now return a spread of values (and a note that they almost hit saturation in a third of the full exposure time).
This is a better design that returns a big picture with more "actionable infomation" - and it's almost "pre algorithm" raw information as it's likely that sensor data and sub region exposure time meta data would be available via a custom RAW dump.
I agree with you, perhaps more fully than you intend, that all hardware level design decisions make up the first (or perhaps 0'th) level of 'back-end processing'.
I also approve of the approach here, one large sensor array is now many small independent sensor regions.
Done properly it becomes and improvement, much as James Webb having many small individually tunable receptor regions was one of many improvements over prior space telescopes.