Myth busted: small pixels have worse performance than large pixels

[Daniel Browning]

But I was also saying that in the end it doesn't matter because of the progression and evolution of the technology. While one photosite design and manufacture process may give diminishing returns if scaled to a smaller size, that does not necessarily hold true for the next design.

Agreed. I think I see your point now; thank you for the clarification.

We're beneficiaries of the ingenuity of sensor designers.

Also not sure how you determine there is less than a full stop of improvement.

Older sensors had QE as low as 25%, so back then it was theoretically possible to get a two stop improvement (log(100/25)/log(2)). Now that many sensors are easily up to 50%, it would require going from 50% to 100% in order to get one full stop (log(100/50)/log(2)) of improvement to QE.

Close is subjective. To me, a 50-70% range still represents a long way to go, especially when that last 30% will be the most difficult.

Agreed; it is a very long way to go because of how difficult that is, but the maximum improvement is only 30%, whereas the other two important aspects of sensor performance, read noise and dynamic range, have no theoretical limit on improvement. QE has slowed more and more as we've gotten closer to the theoretical limit in QE. Read noise, on the other hand, has improved steadily, and there are plenty of papers and patents on dynamic range improvements (though they may have some issues since they've not been commercially utilized except at Fuji). Although there appears to be many practical limitations to improving read noise and DR, I think sensor designers will find a way somehow.

But as for QE, although 100% is theoretically possible, I don't think it's practical in the foreseeable future, since it would require removing the color filters, and therefore a new method to detect color. (One possibility would be directly measuring the frequency and phase of light itself, but that's in the realm of sci-fi.)

 

[Steven Caesare]

Marketing snake oil... been with photography for a long time I think.

IBloom

Certianly true in marketing, etc... but I meant from people who should know better, 3rd parties doing shootouts, etc...

 

[jbeale]

Wow! I'm still absorbing all this information, but it looks like my "smaller sensors = worse performance, other things being equal" idea is well debunked here. I'd seen the original dpreview comparison claiming that downscaling did not work well, but missed the forum posting showing that when done correctly (from RAW), in fact it worked very closely to theory.

One thing that demonstrates to me is that (for the 50D anyway) the readout/ADC noise is really negligible; otherwise you would see more of a performance hit from having more photosites on the same overall area. All very interesting.

The mention of 90% QE at some wavelengths is amazing. Is that on a consumer device? So extreme measures like backside thinning seem not worth the while. Or maybe that 90% is only at the far red end of the spectrum.

(...I'll be happy to post additional provocative and wrong conventional wisdom here, if it gets responses as illuminating as this one!)

 

[jbeale]

more about Small vs. Big pixels

more about Small vs. Big pixels

The most careful and detailed exposition on digital camera image quality I have seen online is from Emil Martinec at the University of Chicago. Daniel, I wonder if you would agree with his comment below regarding pixel size:

"There is an advantage to big pixels in low light (high ISO) applications, where read noise is an important detractor from image quality, and big pixels currently have lower read noise than aggregations of small pixels of equal area. For low ISO applications, the situation is reversed in current implementations -- if anything, smaller pixels perform somewhat better in terms of S/N ratio (while offering more resolution)."
-Emil Martinec, http://theory.uchicago.edu/~ejm/pix/20d/tests/noise/noise-p3.html

 

[Daniel Browning]

Thanks, John.

The mention of 90% QE at some wavelengths is amazing. Is that on a consumer device?

Some of the SBIG monochrome sensors are very close. The consumer sensors tend to come with their CFA intact, but even so they're already hitting 70% at some wavelengths; with no filter they can hit 90%. Unfortunately none of the manufacturers provide that data, and it takes a monochromator to measure. Some of the sensor specs provide the wavelength along with the e-/pixel/lux second specification, which could be converted to QE using the pixel size if one had the desire.

So extreme measures like backside thinning seem not worth the while.

That's what some sensor designers are wondering, such as the guy that runs the Image Sensors World blog. But I think there will be enough improvements to other areas such as FWC and angle of response to make it worthwhile, now that BSI can be done so inexpensively.

Daniel, I wonder if you would agree with his comment below regarding pixel size:

It's pretty similar to what I wrote above on "Read Noise", but mine is much more understated.

There is a certain sensor design that reduces read noise to very low levels, but it only happens to occur in some sensors with big pixels (4.7+ microns), analog gain, and is accompanied by high read noise at low gain (compared to small pixels). Not all big pixels have this design (e.g. RED ONE), and not all big pixels with analog gain have this characteristic either. Yet for every decrease in pixel size, this design continues to have effect.

So the question then becomes: why doesn't this design and characterstic appear in very small pixel sensors? I think the most likely reason is that it has a natural correlation with pixel size, or perhaps pixel size for a given semiconductor process. But what size and/or process? Since large pixels only tend to exist in large sensors, and pixel size in large sensors is limited by processor and storage (rather than sensor design), we can't know how small the pixels can be before this design stops working.

Since history has shown the large pixel high gain designs to scale perfectly as they shrink smaller and smaller, we can't yet say where the correlation is between pixel size and this characteristic except that they haven't hit any limits yet.

 

[Michael Brennan]

A comment about what I understand about noise.

A single large pixel receives more hits from photons than a smaller pixel.
Processing noise from A/D and amps seems to be problem regardless of the size of the pixel. Thermal noise is also a minor issue.
(Astronomers, who are trying to differentiate between a few photons from a distant planet and camera noise, go to the trouble of cooling a ccd to reduce thermal noise which is more a problem for them due to long duration exposures)

Electronic noise caused by processing and thermal noise create random increases in output (noise). lets say all ccds processing has the equivalent of one photon of random noise per pixel. This level has little impact on a large pixel where the darker areas of a scene are being impacted by say 4 photons.

But the smaller pixel may only be receiving 2 photons from the scene, so the one photon level of noise is more obvious.

As I understand it, refocusing an image from 2/3 to 1/3 does not mean twice as many photons hit the smaller pixel.

So all things being equal large pixels and relatively low camera noise enable greater use of shadow areas and hence greater usable dynamic range.


However this is a moving target as noise reduction in processing is improving.

S/N ratio is a useful reference to compare noise between electronic cameras.


Mike Brennan

 

[Daniel Browning]

lets say all ccds processing has the equivalent of one photon of random noise per pixel. This level has little impact on a large pixel where the darker areas of a scene are being impacted by say 4 photons.

But the smaller pixel may only be receiving 2 photons from the scene, so the one photon level of noise is more obvious.

That position is based on the assumption that all pixel sizes have the same read noise, but in reality they do not. For example, at base amplification, there are 2 micron pixels (LX3) that have a read noise equivalent to 5.6 photons, and 6.4 micron pixels (5D2) that have a read noise equivalent to 23.5 photons, as I pointed out above.

As I understand it, refocusing an image from 2/3 to 1/3 does not mean twice as many photons hit the smaller pixel.

What does "refocusing an image from 2/3 to 1/3" mean?

 

[Nova Invicta]

The next can of worms

The next can of worms

Image sensors and their processing of images are equally initially affected by the taking lens, its focal lenth, its stop, CA, field curvature, fringing etc etc. To date most lenses for cinematography have been biased towards film and not to different characteristics for CMOS or CCD imagers so the contrast ratios have not been ideal. In an ideal world lenses and sensors would be optimised to each other with software used to iron out slight variances as zooms for instance zoomed in or out.

 

[Michael Brennan]

The clear fact remains that big size pixels are less noisy than smaller size pixels.
Noise is a hugh bugbear for DPs!


Michael Brennan

 

[Steven Caesare]

Did you READ this thread?

 

[Michael Brennan]

Steven,

Yes.

What is the point of your question and why did you make it in capitals?



Michael Brennan

 

[Steven Caesare]

Because after Daniel spent significant time and energy actusaly describing why this isn't a supported premise, your post is the logical equivalent of "nuh-uh".

Perhaps if you are going to state a contrarian view on a thread Daniel started, you might care to make a cogent counter-argument and maybe even provide some (gasp!) facts or references so we aren't just down to "because you said so"?

-Steve

 

[DS Williams]

Mr. Browning.

HOW THE HELL DO YOU KNOW SO MUCH.

I greatly respect you. Seriously man you're always around to clarify or straighten things out. Much of this is over my head but I'll research it and try to comprehend it.

Thank you for this wonderful thread, as well as your other one.


- DW

 

[Michael Brennan]

References

References

.

Perhaps if you are going to state a contrarian view on a thread Daniel started, you might care to make a cogent counter-argument and maybe even provide some (gasp!) facts or references so we aren't just down to "because you said so"?

-Steve

Steve,

You are getting wrong end of the stick, I don't think I have contradicted any part of Daniel's post, I am re-stating the fact that big pixels are less noisier than small pixels.
Why? I personally believe this is a more important issue for DPs given high compression in camera and distribution than the benefits of smaller pixels.
Most of the research work revolves around analysis of 1st generation still photography images which does not involve lossy compression of a motion picture or TV workflow which exacerbates noise.

This puzzle could be described as "noise versus resolving power".

If you want references of the fact that bigger pixels are less noisy here are some of them, a few going into similar detail and conclusion as Daniel's post comparing small and large pixels.


Here is a selection;


From Emil Martinec professor of physics university of chicago

Noise, Dynamic Range and Bit Depth in Digital SLRs -- Supplemental page 3a
So in comparing large pixels to small pixels, if the issue is how well are the pixels performing, the comparison should be made at high ISO where the measured read noises of each are due to the contributions from the photosite. By this measure, the big pixels always win when data from current cameras are compared; not because they collect less light -- as was seen on page 3, the light collection efficiencies are comparable -- but because the read noise of the larger pixels is smaller on a per area basis.


DXO Labs
"For a given exposure time, each smaller pixel receives four times less light than a large pixel—the equivalent of reducing exposure time by a factor of four. So to get the same sensor response, exposure time needs to be multiplied by four, which means that the ISO sensitivity of the high-resolution sensor is four times less than the low-resolution sensor.
Now assume that the same exposure times and identical ISO settings are used with a low-resolution camera and with a high-resolution camera having four times as many pixels. Since each high-resolution pixel is intrinsically less sensitive, a higher gain (either analog or digital) is applied to the signal, yielding more noise."

Lars kjellberg
QPcard - Dynamic range and pixel size
"Big pixels will give you cleaner, smoother*- less noisy*- shadow areas."


R N Clark
The maximum number of photons one can count with a sensor is the max number of electrons in that can be held in the well. There is one electron per photon. If one fills the pixel well with 40,000 electrons, then the noise in the signal is square root 40,000. So whatever the signal is, the error (noise) is square root of the number of electrons (photons). The more photons counted, the higher the signal-to-noise. The signal-to-noise = # photons/square root(# photons) = square root(# photons) In the shadows in an image, one may get only a few hundred photons, so the noise is square root of those few hundred.
Larger pixels have reduced noise at all levels, but especially at low signal levels
Because good digital cameras are photon noise limited, the larger pixels will always have higher signal-to-noise ratios unless someone finds a way around the laws of physics, which is highly unlikely.

David Frosini
Application Engineer Kodak
"pixel area also affects the dynamic range of the camera system. Each pixel is like a well that stores up charge during its exposure period before it is read out. The larger the pixel area the larger the well capacity, and the larger the charge capacity and the higher the signal level. A pixel is saturated when further increases in illumination do not relate a corresponding increase in signal voltage. Larger pixel areas help improve dynamic range because they hold more charge, so brighter objects do not saturate the pixel. In fact the technical definition of dynamic range is the ratio of the saturation voltage to the total noise of the DSC. Because camera system noise is generally independent of pixel pixel area, increasing the pixel size gives a higher S/N ratio.


Ting Chen, Peter Catrysse, Abbas EL Gamal, Brian Wandell
Information Systems Laboratory Stanford University.
The work reported in this paper was supported under the Programmable Digital Camera Project by Intel,Hewlett-Parkard, Kodak, Interval Research, and Canon.

"Pixel design is a key part of image sensor design. After deciding on pixel architecture, a fundamental tradeoffs made to select pixel size. A small pixel size is desirable because it results in a smaller die size and/or higher spatial resolution; a large pixel size is desirable because it results in higher dynamic range and signal-to-noise ratio. Given these two ways to improve image quality and given a set of process and imaging constraints an optimal pixel size exists. It is difficult, however, to analytically determine the optimal pixel size, because the choice depends on many factors, including the sensor parameters, imaging optics and the human perception of image quality.

In summary a small pixel size is desirable because because it results in higher spatial resolution and better MTF. A large pixel size is desirable because it results in better DR and SNR"
http://isl.stanford.edu/~abbas/group/papers_and_pub/pixelsize.pdf

 

[Jannard]

Just for fun... how do you explain the fact that one sensor vastly outperforms another when they have exactly the same pixel size and are on the same size sensor? Pixel size alone tells you very little. Pixel design along with pixel size tells you a lot more. For example, the Mysterium sensor and Mysterium-X sensor both have 5.4 micron pixels, yet the X sensor is significantly better in performance. Why? Pixel design. The biggest breakthroughs in sensor design are coming from pixel design advancements.

Jim

 

[Von Thomas]

I wish I had the digital files, but I don't. I compared two digital backs from a well known company, one back 39MP, the other 22MP. Both backs same size sensor, but the 39MP has pixel size of 6.8 microns, and the 22MP has pixel size of 9 microns.

I ran tests on both to see if there was any difference in the files, and I found that the 39MP back displayed higher noise in the shadow areas of a moody light scene, when compared to the 22MP. I found not only noise in the transitions from highlight to shadow with the higher MP back, but also ugly artifacting.

This was done with same camera body, same lens, same exposure settings, just swapped the backs.

That proved to me that the larger MP, smaller photo site sensors were not as clean as their larger pixel cousins.

I think to be sure, maybe someone can grab a Canon 1DS MKII, and a 1DS, and possibly a 1DS MKIII, and let's have a real test.

 

[Steven Caesare]

I think the point is that the "bigger-pixels = better" statement as made in a vacuum, doesn't necessarily translate to "worse images", which is the crux of Daniel's point.

 

[Alan Fletcher]

Good article and thread. :hurray:

Steve,

DXO Labs
"For a given exposure time, each smaller pixel receives four times less light than a large pixel—the equivalent of reducing exposure time by a factor of four. So to get the same sensor response, exposure time needs to be multiplied by four, which means that the ISO sensitivity of the high-resolution sensor is four times less than the low-resolution sensor.
Now assume that the same exposure times and identical ISO settings are used with a low-resolution camera and with a high-resolution camera having four times as many pixels. Since each high-resolution pixel is intrinsically less sensitive, a higher gain (either analog or digital) is applied to the signal, yielding more noise."

That's just plain wrong. Given the SAME technology the small and big pixels have the same photon flux (electrons) per unit area, and produce the same VOLTAGE (based on the above-given capacitance formula V=Q/C, where C is proportional to Area). They therefore have the same Base ISO.

The difference between them (in isolation) is that the smaller sensor will have fewer electrons, and therefore higher SHOT noise.

I have been somewhat convinced by posts on dpreview that for GOOD light (highlights, full wells) you probably gain more from the extra spatial information than you lose from the extra noise.

But for low light (high ISO or shadows, nearly empty wells) I am still convinced that big pixels are better.

That's not to say that ever-evolving good pixel design doesn't make the next generation better than the last, or that different design trade offs are made for small vs big pixels.

ps -- I base my signal-to-noise anaylsis on Janesick's 2002 paper, where he gives the signal-to-noise equations for an ideal camera. Unfortunately my link to it has rotted.

pps -- There's no point adding a 18-bit DAC to the chain when all it will do for you is count each electron TWICE.

[ edit : removed repeated word]

 

[Chris Newman]

I ran tests on both to see if there was any difference in the files, and I found that the 39MP back displayed higher noise in the shadow areas of a moody light scene, when compared to the 22MP. I found not only noise in the transitions from highlight to shadow with the higher MP back, but also ugly artifacting.

Did you resize the images to be the same size that is smaller than the lower-res sensor? This is an important requirement for Daniel's argument, and one that I think people commonly overlook.

 

[Jeff Kilgroe]

A big thanks to Jim for stating so clearly what I was unable to convey in my much lengthier post. Pixel size tells us very little, it's about pixel design and how each photosite on a sensor works to create the entire sensor surface.

Daniel is right though, the typical reasoning that larger pixels equate to a better image is largely a myth. However, it's difficult to quantify that because there are very few real-world examples where sensors with different sizes of photosites will still share equivalent pixel designs. In most cases, comparing one sensor to another sensor from the same manufacturer, even within the same product line, we are almost always comparing different designs and generations of sensors. By and large, in the sensor market, as resolution increases, so do other capabilities of the sensors as resolution and light gathering ability are continuously evolved.

Same thing with dynamic range and noise. This is more a factor of pixel design and not pixel size. Large photosites do not equate to less noise as many people often believe. In fact, it's often quite the opposite. Additionally, if I were looking to dig more usable detail out of a noisy image, increasing the resolution by way of more photosites on a sensor is one of the best ways to preserve fine details in high-noise conditions.

It will be interesting to see how the 5.4 micron Mysterium-X compares with the 3 microng Mysterium-X. While they share the same namesake, I doubt they share the same pixel design. I expect the sensor with smaller pixels to perform equally well as the larger version in terms of image quality. The pixel design, being smaller, will most likely make it's compromises in other areas such as maximum speed, reset rates, etc.. But that's pure speculation.