What happened to "CCD is better than CMOS"?!

[BobShaw]

Motorbikes are better than trucks. It depends who you ask and what they are using them for.

IMHO my 8 year old Hasselblad with CCD has far better IQ than any 35mm DSLR with CMOS for studio and landscape. It also has ISO 50, which I prefer. However if I am going to sports or low light then it will be the Canon.

BTW the IBM Power PC chip was the chip used in the IBM Mini computer that a lot of medium size enterprises used. It became a case of cost versus benefit and connectivity, and IBM moved at the speed of, well IBM.

[ErikKaffehr]

Hi,

That may depend on definition of "Pipeline". A basic difference between CCD and CMOS is that with CMOS it is possible to make multiple readouts. A CMOS sensor can be read out after reset. So the camera can record electron charges on all pixels before exposure and this reading is subtracted from the electron charges after exposure. This is called "correlated double sampling" and is always done on CMOS, AFAIK. CCD cannot do this, as CCD readout is destructive. All electron charges are removed from a CCD during readout.

All CMOS is not alike. Modern CMOS sensors, like those made by Sony, use massively parallel ADCs. The readout is a part of the sensor, there is an ADC (Analogue Digital Converter) for each column. Those converters are very simple, but accurate, ramp type converters.

Canon uses a different approach, they have off sensor ADCs and few readout channels. So they have a longer signal path and need much faster converters that are more complex. That is the reason that Sony sensors have better DR at base ISO than Canon sensors. Sony is not the inventor of massively parallell readout, there are other companies using the technology.

With CCD there is always a dark frame subtraction, CMOS doesn't really need it at short exposures.

Comparing CCD with CMOS the analogue processing is done by Sony, and the signal coming from the sensor is digital.

But, all things we have discussed here are part of the electronic processing pipeline. The end of this pipeline is the raw image. The raw image would ideally be just a dump of the recorded voltages and complemented recorded information. But raw files are probably not really raw. Just as an example the raw data is often loss-lessly compressed.

The image part of the raw file is just numbers. No difference between CMOS and CCD. The raw file also contains metadata and that is vendor specific. Here are some examples of EXIF data: http://echophoto.dnsalias.net/ekr/Articles/EXIFSample/

From here on the pipeline would ideally be the same, but it is my understanding that CCD processing includes additional steps. CCDs have often a "tiling issue", that needs to be corrected in the processing pipeline.

Large beam angles on technical cameras, or biogon type lens designs on mirrorless can cause "crosstalk" those issues are present on both CCD and CMOS, but at least Sony's CMOS designs are more sensitive to beam angles than "Kodak" CCDs used older devices.

My understanding is that much more corrections are needed to CCD files than to CMOS files. This is a comment from Anders Torger who has down a lot of work in raw conversion of CCD images: http://forum.luminous-landscape.com/index.php?topic=106969.msg882144#msg882144

Once the initial stage of processing is done the pipeline would be shared between CMOS and CCD. Demosaic is done, that is the missing values in each RGBG quartet is interpolated from surrounding pixels. Keep in mind that pixels are monochrome. But there is a colour filter in front of each pixel. After demosaic, all pixels have three signals, one measured and two interpolated. These signals are still just numbers.

Next stage, each pixel (that has three numerical values) is multiplied with a colour conversion matrix that converts it's three numbers to a colour space, more often than not XYZ. So it is this conversion matrix that supplies the colour. Such a matrix is just 9 numbers in 3x3 pattern. It could be said that colour transformation matrix is a description of the Color Filter Array in front of the sensor. I have attached colour sensivity curves measured on two digital backs one from Phase One the other from Hasselblad. The source was not more specific on which models. A customer camera is also include. Those curves are represented by a simple matrix having nine numbers.

There are a lot of steps ahead in the processing pipeline. White balancing being one of the most important.

Best regards
Erik

Back to topic with an innocent question, continuing on mjrichardson stream of thoughts:

Is the pipeline behind CCD and CMOS completely interchangeable? If not then we can surely talk about color reproduction of CCD-sensor-stack and CMOS-sensor-stack?

[ErikKaffehr]

Hi,

Intel is very good at chip making. They are doing a lot of continuous improvement and they have the money and scale of economy so they can stay on the forefront of making processors. Many of the inventions are coming from other places, like AMD. But once Intel acquires a technology they are extremely good in refining it.

Intel is also a great master of marketing tricks and they can afford expensive lawyers.

IBM on the other side is more like a system integrator than a high volume chip maker.

Best regards
Erik

Motorbikes are better than trucks. It depends who you ask and what they are using them for.

IMHO my 8 year old Hasselblad with CCD has far better IQ than any 35mm DSLR with CMOS for studio and landscape. It also has ISO 50, which I prefer. However if I am going to sports or low light then it will be the Canon.

BTW the IBM Power PC chip was the chip used in the IBM Mini computer that a lot of medium size enterprises used. It became a case of cost versus benefit and connectivity, and IBM moved at the speed of, well IBM.

[hjulenissen]

Hi,

That may depend on definition of "Pipeline". A basic difference between CCD and CMOS is that with CMOS it is possible to make multiple readouts. A CMOS sensor can be read out after reset. So the camera can record electron charges on all pixels before exposure and this reading is subtracted from the electron charges after exposure. This is called "correlated double sampling" and is always done on CMOS, AFAIK. CCD cannot do this, as CCD readout is destructive. All electron charges are removed from a CCD during readout.

All CMOS is not alike. Modern CMOS sensors, like those made by Sony, use massively parallel ADCs. The readout is a part of the sensor, there is an amplifier for each column. Those converters are very simple, but accurate, ramp type converters.

Canon uses a different approach, they have off sensor ADCs and few readout channels. So they have a longer signal path and need much faster converters that are more complex. That is the reason that Sony sensors have better DR at base ISO than Canon sensors. Sony is not the inventor of massively parallell readout, there are other companies using the technology.

With CCD there is always a dark frame subtraction, CMOS doesn't really need it at short exposures.

Comparing CCD with CMOS the analogue processing is done by Sony, and the signal coming from the sensor is digital.

But, all things we have discussed here are part of the electronic processing pipeline. The end of this pipeline is the raw image. The raw image would ideally be just a dump of the recorded voltages and complemented recorded information. But raw files are probably not really raw. Just as an example the raw data is often loss-lessly compressed.

The image part of the raw file is just numbers. No difference between CMOS and CCD. The raw file also contains metadata and that is vendor specific. Here are some examples of EXIF data: http://echophoto.dnsalias.net/ekr/Articles/EXIFSample/

From here on the pipeline would ideally be the same, but it is my understanding that CCD processing includes additional steps. CCDs have often a "tiling issue", that needs to be corrected in the processing pipeline.

Thank you for a nice overview.

I would not worry about lossless compression (as it does not change the information once it is properly decoded). But what about sensor non-uniformity, individual "outlier" sensels, etc? It would be very interesting to get to know what kind of processing manufacturers does between the ADC and storing a raw file. Some kinds of processing might be hard to detect (e.g. median-type filtering), while others should be quite detactable (multiplying 14-bit numbers by a per-pixel static gain array in the range of [0.5...2.0] and storing the result as non-dithered 14 bits). The complexity of doing things in the CFA domain along with the required pixel rate and battery life is probably limiting what the manufacturers can do.

The fact that camera manufacture is competitive and people are (to some degree) basing purchases on comparing raw files (or raw file developers) makes it tempting to make the raw files "shine" even if that does not contribute to the final developed image quality (even if it detracts slightly).

-h

[ErikKaffehr]

Hi,

My understanding is that median filtering is very easily detectable. You just shoot a dark exposure and do FFT. The resulting image would show a drop off to high frequencies.
Jim Kasson has several very nice articles demonstrating this. Very clearly Sony uses this at very high ISO settings.

This is illustrated in the enclosed screen dump from Bill Claffs PDR page: http://www.photonstophotos.net/Charts/PDR.htm

The curve has two bumps, the first at 640 ISO (nominal) the other one at 25600 ISO (nominal). The first one (at 640 ISO) is the effect of a trick invented by Aptina, the second one at 25600 ISO is the effect of median filtering.

I am not concerned about lossless compression, or lossy one for that part. But I feel it should be mentioned in the context.

Personally, I guess that a raw file should be very little processed.

Best regards
Erik

Thank you for a nice overview.

I would not worry about lossless compression (as it does not change the information once it is properly decoded). But what about sensor non-uniformity, individual "outlier" sensels, etc? It would be very interesting to get to know what kind of processing manufacturers does between the ADC and storing a raw file. Some kinds of processing might be hard to detect (e.g. median-type filtering), while others should be quite detactable (multiplying 14-bit numbers by a per-pixel static gain array in the range of [0.5...2.0] and storing the result as non-dithered 14 bits). The complexity of doing things in the CFA domain along with the required pixel rate and battery life is probably limiting what the manufacturers can do.

The fact that camera manufacture is competitive and people are (to some degree) basing purchases on comparing raw files (or raw file developers) makes it tempting to make the raw files "shine" even if that does not contribute to the final developed image quality (even if it detracts slightly).

-h

[mjrichardson]

Morning

As a non techie person who takes photographs the answer to this very simple question is that if you accept that the camera produces a file, however that is done internally, then it is easy to decide which you prefer based on how it works for you. There is no better or worse, there is better for you or worse for you. I am very comfortable in saying that when I take a portrait with both my cameras at base ISO, I always prefer the output from the camera which uses a ccd sensor, that's a fact for me. If I need to shoot at ISO 800 then I always prefer the output from the camera using the cmos sensor but it doesn't match the quality of the ccd at base ISO obviously but beats it significantly at ISO800.

We can all express a personal preference, we can pedantically state that it is electronic device and as silicon has no inherent look but this is the real world, we buy a camera that is a finished product and we use it to create our own finished product and can find that our choice of camera is different to everyone elses and there is nothing wrong with that, we all work differently. A camera that provides advances in areas that you have no need for in your work is not better if it doesn't provide advantages for the way you do work, surely?

Have a good day.

Mat

[LKaven]

Intel is very good at chip making. They are doing a lot of continuous improvement and they have the money and scale of economy so they can stay on the forefront of making processors. 

The i86 was one of the most inefficient designs ever for a microprocessor.  [They never heard of a register file apparently, and this mistake would cost them billions.]  It would have passed into history ironically if IBM itself had not made it the center of their first PC.  The MC68k was a much cleaner design.  The PowerPC chip, given the same amount of refinement, could have been more competitive.  I don't think today's ARMs are that much different.

[synn]

What happened to "CCD is better than CMOS?

This exact same buncha posts happened, have been happening for over a decade now.

[ErikKaffehr]

Hi,

This is a technical thread on a technical subforum ("Equipment & Techniques"), so what would you expect?

What has changed now is that Phase One has in part switched to Sony CMOS for it's top of the line camera. I know that you are perfectly happy owner of a CCD based Leaf camera and I am pretty sure it can make great images.

Actually, almost any camera can make great pictures and no sensor will make any a better photographer. There are some techniques that some cameras make possible like tilt and movements, or short exposures using strobe light. But that is a camera thing and not a sensor thing.

What a new sensor can give you is more resolution, better DR, less noise and less aliasing. All these are incremental things, but as all those things come at a significant I would suggest that it can be worthwhile to discuss benefits.

Switching to CMOS doesn't necessarily mean Phase One would abandon CCD, it depends on who well they sell, but I think they will be less enthusiastic on the advantages of CCD.

Best regards
Erik

This exact same buncha posts happened, have been happening for over a decade now.

[ErikKaffehr]

Hi,

Yes I agree on that. I used to be a MC68K advocate before Sun introduced SPARC. I can also agree that the Intel X86 architecture is ugly.

But intel has been very successful in pushing the envelope of that ugly architecture.

Best regards
Erik

The i86 was one of the most inefficient designs ever for a microprocessor.  [They never heard of a register file apparently, and this mistake would cost them billions.]  It would have passed into history ironically if IBM itself had not made it the center of their first PC.  The MC68k was a much cleaner design.  The PowerPC chip, given the same amount of refinement, could have been more competitive.  I don't think today's ARMs are that much different.

[hjulenissen]

Hi,

Yes I agree on that. I used to be a MC68K advocate before Sun introduced SPARC. I can also agree that the Intel X86 architecture is ugly.

But intel has been very successful in pushing the envelope of that ugly architecture.

If it was practically (i.e. with some ease) possible to make something 2x or 10x more efficient (cost, power, area,...) without sacrificing compiler complexity, programmer effort etc, then surely we would have seen it by now?

After all, process tech in the free market is only slightly behind Intel now. So a cpu startup could go from nothing to designing and verifying a design, producing them at some silicon cooking plant, then drive Intel out of business. Provided that Intels design/instruction set is such a limitation. Previously, the Wintel monopoly has been presented as the reason why this won't happen, but now we see smartphones and tablets and servers and embedded boxes using non-Intel and non-x86 processors.

It is possible that through some strike of luck, genious or massive effort, someone designs something that is vastly better than Intel in all of those criteria, but I won't hold my breath.

The industry is full of claims about 10x this and 10x that, but upon closer inspection, it seems that they either 1)Never get a product out, or 2)optimize e.g. performance vs power by sacrificing generality and programmer effort (i.e. GPU).

-h

[ErikKaffehr]

Hi,

An interesting discussion but pretty much off thread…

Best regards
Erik

If it was practically (i.e. with some ease) possible to make something 2x or 10x more efficient (cost, power, area,...) without sacrificing compiler complexity, programmer effort etc, then surely we would have seen it by now?

After all, process tech in the free market is only slightly behind Intel now. So a cpu startup could go from nothing to designing and verifying a design, producing them at some silicon cooking plant, then drive Intel out of business. Provided that Intels design/instruction set is such a limitation. Previously, the Wintel monopoly has been presented as the reason why this won't happen, but now we see smartphones and tablets and servers and embedded boxes using non-Intel and non-x86 processors.

It is possible that through some strike of luck, genious or massive effort, someone designs something that is vastly better than Intel in all of those criteria, but I won't hold my breath.

The industry is full of claims about 10x this and 10x that, but upon closer inspection, it seems that they either 1)Never get a product out, or 2)optimize e.g. performance vs power by sacrificing generality and programmer effort (i.e. GPU).

-h

[torger]

Well, it won't matter in a few years anyway, because CCD is with all likelihood going away.

We could argue about CCD vs CMOS when CCD=MFD and CMOS=135, but this is no longer true. MFD is CMOS too now and CCDs are being phased out.

[AlterEgo]

But intel has been very successful in pushing the envelope of that ugly architecture.

because it is not longer the original "x86" under the hood...

[Theodoros]

Well, it won't matter in a few years anyway, because CCD is with all likelihood going away.

We could argue about CCD vs CMOS when CCD=MFD and CMOS=135, but this is no longer true. MFD is CMOS too now and CCDs are being phased out.

I too believe that there will be no new CCD sensors coming, but IMO this isn't because of IQ as I believe many are happy with the "looks" of CCD images and don't care much about higher ISO as they would use a DSLR for that anyway... IMO the main reason on why there will be victory for Cmos sensors is their LV performance... Never the less, the few lucky ones that use multishot backs for most part of their work, won't even bother to check if the next (again multishot) back is CCD or Cmos (since the phrases "presence of artifacts", "enough resolution" & "color presantation" are unknown to them)... but again, LV will also matter to them too...

[ErikKaffehr]

Hi,

I guess that for those who need LV, me among them, a well working LV system is very important.

Regarding ISO speed, it can be clearly beneficial to have good high ISO when shooting outdoors, say in windy conditions.

Regarding "multishot" it is great for sure in a studio, but again I would say it is a great advantage to be able to capture with a single exposure.

Best regards
Erik

I too believe that there will be no new CCD sensors coming, but IMO this isn't because of IQ as I believe many are happy with the "looks" of CCD images and don't care much about higher ISO as they would use a DSLR for that anyway... IMO the main reason on why there will be victory for Cmos sensors is their LV performance... Never the less, the few lucky ones that use multishot backs for most part of their work, won't even bother to check if the next (again multishot) back is CCD or Cmos (since the phrases "presence of artifacts", "enough resolution" & "color presantation" are unknown to them)... but again, LV will also matter to them too...

[Franzl]

I too believe that there will be no new CCD sensors coming, but IMO this isn't because of IQ as I believe many are happy with the "looks" of CCD images and don't care much about higher ISO as they would use a DSLR for that anyway... IMO the main reason on why there will be victory for Cmos sensors is their LV performance... Never the less, the few lucky ones that use multishot backs for most part of their work, won't even bother to check if the next (again multishot) back is CCD or Cmos (since the phrases "presence of artifacts", "enough resolution" & "color presantation" are unknown to them)... but again, LV will also matter to them too...

Wasn't ccd dead already and had a come back and now cmos is here and who knows what technology comes out regarding ccd. Doesn't matter as long as there is a pic coming out of that cam which gives you the pixels you need to achieve the photo you want.

Just wanted to mention that it is nonsense that you only need high ISO in 35mm format. I often needed DMF and ISO400, because the client would make a billboard and crop out a city light or an ad from that horizontal shot. And sometimes you cannot flash everything to not kill the mood and you need some of that ambient light. ISO400 or iso800 for portrait jobs are often needed. And a DSLR will always look like a DSLR and won't have the look of DMF. So the higher ISO are more than welcome and were the reason I upgraded from P45+ to IQ150. CMOS is finally here in DMF and if this means higher ISO, I am thankful for that and don't care if it is due CCD or CMOS.

[bjanes]

Hi,

I guess that for those who need LV, me among them, a well working LV system is very important.

Regarding ISO speed, it can be clearly beneficial to have good high ISO when shooting outdoors, say in windy conditions.

I agree that Live View is of high importance in some situations, and a well implemented electronic first curtain shutter (not available with CCD) is important when photographing at high magnification with a long telephoto or microscope. I say well implemented because the implementation on the Nikon D810 is suboptimal.

Regards,

Bill

Bill

[ErikKaffehr]

Hi Bill,

Agree on all points. I don't own a Nikon, so I cannot comment on D810, of course.

I would also like to see an articulated LCD, it is so helpful when shooting at low position.

Best regards
Erik

I agree that Live View is of high importance in some situations, and a well implemented electronic first curtain shutter (not available with CCD) is important when photographing at high magnification with a long telephoto or microscope. I say well implemented because the implementation on the Nikon D810 is suboptimal.

Regards,

Bill

Bill

[BernardLanguillier]

I agree that Live View is of high importance in some situations, and a well implemented electronic first curtain shutter (not available with CCD) is important when photographing at high magnification with a long telephoto or microscope. I say well implemented because the implementation on the Nikon D810 is suboptimal.

Bill,

Are you referring to the fact that MLU is needed to use EFC on the D810?

Cheers,
Bernard