DSLR CMOS vs CCD BAYER SENSOR ANALYSIS


DSLR CMOS vs CCD BAYER SENSOR ANALYSIS



DSLR CMOS vs CCD BAYER SENSOR ANALYSIS






Read any article about CCD vs CMOS imaging sensors and you will be informed quite categorically that whereas the maturing CMOS architecture developed for DSLR cameras is good enough for everyday photography, when it comes to astro-imaging the CCD still reigns supreme. (<http://www.dalsa.com/corp/markets/ccd_vs_cmos.aspx> on-line ref chosen at random - there are thousands of such articles, all basically telling you the same message, just Google CCD vs CMOS & see for yourself).

Whether this argument can be sustained in practice when it comes to commercially available DSLR's however is another matter entirely. Because, "everybody knows", CCD sensors are less noisy, have higher quantum efficiency and dynamic range, and are to be found almost exclusively in astro-imaging cameras, there is an unstated inference this also applies to DSLR cameras.

CMOS sensors are supposedly inherently more noisy than CCD sensors. Since most DSLR's use CMOS sensors, exactly how noisy are they, and which DSLR's are the least noisy? Also are DSLR's with CCD sensors significantly less noisy than their CMOS cousins?

To address these questions I extracted data on a wide range of DSLR's manufactured between 2000 & 2009, from Dx0 Labs website <http://www.dxomark.com/> DxO publish sensor data from RAW DSLR images. Signal to noise ratio @ 18% greyscale; Dynamic Range in EV (exposure value); Tonal Range, & Colour Sensitivity in bit levels. SNR values are expressed in decibels dB, 1EV = 3dB.

My approach was to rank DSLR's in terms of how noisy their CMOS & CCD sensors were at their native ISO sensitivity, and by how well their dynamic range was managed across their ISO range.

DSLR CMOS sensors all work in basically the same way. The electron count is read out via multiple gates, using either 12bit or 14bit a/d converters. The basic idea is that noise tends to accumulate in the bottom of the photosite's electron well, so if you want an 8bit tonal range (255 shades of grey or primary colour) then you can happily push the electron count up 4bits without saturating any of the photosites, and delete the bottom 4bits. In theory a 14bit a/d offers you more leeway, but in practice there appears to be no noticeable improvement in tonal or dynamic range.

DSLR CCD sensors use a bucket-brigade electron well readout, typically through 2 gates. CCD's do not have on-chip readout amplification, the readout is analogue and the a/d conversion and any amplification is performed off chip.

Having extracted 8bit colour data per RGB channel plus an 8bit luminance channel, the Bayerised output is demosaiced and converted to a 32bit RAW colour image, comprising 24bit RGB data & 8bit Luminance data.

This all works fine at the CMOS or CCD sensor's native sensitivity, typically ISO100 or ISO200. Problems arise as the ISO rating is increased. The a/d gate readout has to be amplified, and this increases the readout noise. Every 6dB corresponds to a doubling of the readout noise. A typical CMOS or CCD sensor's readout noise will increase by about 3dB for each doubling in ISO rating. 3dB corresponds to 1EV, which means the dynamic range decreases correspondingly. As the readout noise increases the available bit space in either the 12bit, 14bit or in some CCD's 16bit a/d stack gets squeezed. Tonal Range bit depth gets clipped as a way of cutting out noise, and this in turn has an influence on the Colour Sensitivity.

Lets take an example by means of illustration, my Sony alpha900, not because it is a particularly good DSLR noise wise, rather because it illustrates all the drawbacks that come with high photosite density..


On DxO's website, mouse click on each plotted orange dot and you get the plot data:
e.g. SNR@18% @ ISO100 Measured ISO119 SNR 36.3dB & so on.


Using this data I performed a log-linear regression analysis on each variable, and then entered the results into a spreadsheet:

r is the correlation coefficient, the closer it is to 1 the better the log-linear curve fit. Negative r indicates a negative slope, i.e. y values decrease with increasing x values. If r is almost 1 (you will never get r=1 except when sample value n= 2) it informs you, that you have selected the correct function to regress to the mean.

The reason I chose a log-linear function is quite simple to understand. The ISO sensor sensitivity scale is linear, doubling ISO rating doubles sensor sensitivity. The SNR scale is logarithmic. Noise is expressed in dB, each increase by 6dB equates to a doubling of SNR. Likewise Tonal Range & Colour Sensitivity are measured in bits, i.e. to log base 2. I chose log base 10 because converting between base 10 and base 2 is an additional and unnecessary complication.

I analysed the properties of all the Canon, Nikon and Sony stable of CMOS DSLR's using this method, and then ranked the results in order of how well each camera managed its CMOS sensor noise across its ISO range. I also included two CCD sensor Leica rangefinder and several CCD sensor Nikon DSLR's, plus pro-sumer CCD medium format sensor DSLR's.

The ranking method I adopted was based on the SNR@18%. By taking the SNR log-linear relationship, and dividing the constant (where the curve crosses the y SNR axis) by the coefficient of the x logISO value, a notional ISO rating is obtained at which the readout noise will fill the 12bit or 14bit a/d stack.

i.e. logISO = 59 ÷ 10.8 = 5.462 962 63 = K

i.e. DRk = 18 - 3K = 1.611 111 112
i.e. TRk = 12 - 1.8K = 2.166 666 667
i.e. CSk = 32.8 - 5.5K = 2.753 703 705

The overall performance of the sensor is evaluated by summing K + DRk + TRk + CSk and taking the mean:

i.e. RANKING = 11.994 444 45 ÷ 4 = 2.998 611 112 ≈ 3.00 2DP

All the DSLR's analysed were ranked using the same method and listed in rank order:


I think you may well find the results fly in the face of expectations based on, "what everybody knows". Remember this is not ranking according to feel and handling, functionality, build quality, sensor resolution, optical viewfinder & LCD display quality or shooting rate, but CMOS & CCD sensor noise handling across the ISO range. The lower the sensor noise across the ISO range, and particularly at those ISO levels where the full 8bit tonal range is extracted, the higher the ranking.

You can see my Sony alpha900 comes well down the list, but rather surprisingly leading the top of the range Nikon D3X, which uses basically the same Sony Exmor sensor, but with 14bit a/d gates, rather than 12bit a/d gates. Based on noise handling ranking, the top of the range Nikon D3X ranks with the Sony alpha350. This is not to say that the cameras are equivalent. The Nikon D3X has a 24.5Mp Fx CMOS sensor compared to the Sony alpha350 14.1Mp Dx CCD sensor. Images produced by the Nikon D3X will easily out-resolve those produced by the Sony alpha350. But when it comes to deep sky astrophotography, there will be little to choose between them, because that is where low noise, and good noise handling matters most.

Interestingly the DSLR which ranks #1, and by a clear margin from any of the Nikon DSLR's is the Canon EOS 5D which uses a 13.2Mp 12bit a/d Fx CMOS sensor with 8 micron photosites. The closest Nikon offers to this level of performance is its Nikon D3s which is a 12.2Mp 14bit a/d Fx CMOS sensor with 8.4 micron photosites. On the face of it you'd expect the Nikon D3s to have the edge on the Canon 5D because it has slightly bigger photosites and 2bits more a/d stack space to eliminate noise. Bigger photosites equate to higher SNR. But that is not the full story. It is how the camera firmware handles the SNR across the ISO range that is crucial to the overall performance.

A second hand Canon 5D body can be had at present (MAY2010) for about 700. A Nikon D3s retails (body only) for 3600. Bit of a no brainer really.

The other point I want to make about my analysis concerns CCD sensors. It is taken as a given that CCD sensors are inherently less noisy than CMOS sensors. This was the case when the manufacture and architecture of CMOS sensors complied with requirements of logic IC's. Silicon foundries did not have the necessary lithography to manufacture high quality CMOS imaging chips. This is no longer the case, Sony & Canon have silicon foundries capable of producing CMOS imaging sensors to the requisite exacting standard as CCD imaging sensors.

What makes the CCD sensors used in dedicated astro-imaging cameras inherently low noise is a combination of photosite size and well depth, up to 16bit a/d gate, and most importantly cooling. Unfortunately you cannot use a dedicated CCD astro-camera for anything else. Unless you are determined and able to make good use of it, year in, year out, it is dead money.

The virtue of a DSLR is that you can use it for almost any type of photography. It is a versatile portable camera that doesn't need a laptop. But DSLR's do not have cooled sensors, and for print resolution reasons, tend to have photosites ≈ 5 to 7 microns, and readout gates 12bit or 14bit a/d. This makes CCD DSLR's just as noisy, if not more noisy than some CMOS DSLR's.

Lest you doubt this take a good look at the Hasselblad, LEAF, & PHASE ONE medium format DSLR's & the MAMIYA ZD back. I have also included the Leica M8 & M9, although these are rangefinder cameras they have very high quality CCD sensors. None of these cameras have a distinctly less noisy sensor than the majority of CMOS DSLR's. Also note that a 16bit a/d gate does not translate into anything more than an 8bit tonal range, nor does the 8 bit leeway translate into a significantly less noisy image.

DSLR's using Dx format CCD sensors include Nikon D60, Nikon D70, Nikon D80 & Nikon D40X. Whereas they rank in the middle to upper tier, note they are not leading the pack.

Well I hope having perused this article you will be in a position to dispel yet another astro-myth the next time you're at a star party and the dyed-in-the-wool deep sky imagers trot out the same clichéd argument.

If you want a DSLR for every day photography, and as a bonus, capable of deep sky imaging, choose one of the Canon DSLR's from the top of my list. Any DSLR with a ranking 10 or above will do the job very nicely, especially if you have the IR blocker either removed, or a modified IR blocker fitted.

If your DSLR has a lower ranking, like mine, don't worry too much, the differences between the best and the worst are not dramatic. Noise can be reduced by using an external power supply, and leaving a few minutes between exposures, and making sure the noise reduction option is off (in other words take your own separate dark frame). The point I'm making is that noise wise CMOS sensors are every bit as good as CCD sensors in commercially available DSLR's and quite a few are better.

For further reading on DSLR sensor noise and how it affects astrophotography see ClarkVision Digital Camera Sensor Performance Summary

ADDENDUM



Here in the UK there are a few diehard ccd imagers who refuse to accept cmos sensors can have an application in deep sky imaging. Principal amongst them is a character by the name of Grant Privett. Presumably having read this article he felt obliged to e-mail me and acquaint me with "the facts" and offer endless anecdotal evidence, arguing that if cmos sensors were so good why are they not used in deep sky astro-cameras. No amount of hard data would persuade this so-called deep sky imaging guru that simply because astro-camera manufacturers have not as yet used cmos sensors, implies they are inherently noisier than ccd sensors.

It is however a fairly obvious point and one I had already discussed with Terry Platt, the director of Starlight Xpress and the man who started the company in 1992. Terry found my article of interest, and commented that one of the problems astro-camera manufacturers have is Sony's reluctance to make their current cmos imaging sensors more widely available. Sony decline to publish techincal data, or offer their Exmor & R-Exmor sensors to external manufacturers, with the notable exceptions of Nikon and Panasonic. Terry was clearly not of the same ccd diehard school, and informed me he was seriously considering buying a Sony a900 DSLR, dismantling it, and reverse engineering the Fx Exmor sensor, in order to fabricate a prototype cooled cmos astro-camera, despite it being uneconomic.

Privett contributed an article on using a DSLR for astrophotography in the July 2009 issue of Astronomy Now. The article makes curious reading in view of the popularity of DSLR's amongst amateur deep sky astrophotographers. After initially recommending the Canon 20D or Canon's astro-rebel 20Da with the extended IR blocker, he basically argued that if you really want to take deep sky astrophotographs, you'd be better advised getting a dedicated ccd astro-camera. Hardly an enthusiastic endorsement of the article's main aim - i.e. to help and encourage amateurs who own a DSLR to try their hand at deep sky imaging.

When I read the article I found it patronising and condescending. Reading between the lines Grant Privett was posturing as the man who knows about these things because of all the work he has been recently doing in this field, and if you wished to emulate his success, you would be best advised to follow his example. Something struck me as odd about his attitude. It was not clear he had ever used a DSLR for deep sky imaging, but I assume (of course I may be wrong) that if has, then he has managed to obtain some good results. Yet the tenor of his AN article flew in the face of the work of several published authors on the subject, such as Michael Covington, Robert Reeves, Jerry Lodriguss, Ninian Boyle, and several others. It is perhaps significant that the pair of books written by Privett on the subject,"Creating and Enhancing Digital Astro Images" Springer-Verlag 2007,& "The Deep-Sky Observer's Year" Springer-Verlag 2001 - coauthored with Paul Parsons, make no mention of DSLR's.

I formed the distinct impression he was speaking as an idiot savant, a man who has all the mathematical knowledge, yet no idea how to apply it. A case of bias, in asmuch, "Don't confuse me with facts, my mind is made up." So I put some pointed questions to him, insisting on straight answers:

a) I hope, having read my article, you will now accept that CMOS sensors, used in DSLR's are every bit as good noise wise as CCD sensors. If not, I'd like to know what your reasons are for persisting with the claim that CMOS sensors are inherently more noisy than CCD sensors - as used in DSLR's. (In other words lets see a bit of DSLR data not anecdote)

c) If asked to rank the performance of the Nikon D3 or D3s noise wise, would you have ranked it below the Canon 5D? You mentioned the Canon 20D & 20Da in your article. When you wrote it were you aware that it was slightly more noisy than the Canon 5D, based on DxO Labs SNR measures?

d) You comment about 8bit tonal range. When you wrote your DSLR for Astrophotography article did you realise that all DSLR's only write an 8bit RGB RAW image file, no matter what the a/d gate bit depth? (Take a look at DxO Labs tests on Nikon DSLR's - the RAW files output an 8bit tonal range - not 12 or 14bit - 12bit quantized to 9.4bits & clipped - unchanged DR - NEF RAW codec reduces 4096 to 683 - it is not lossless - data is lost, mostly lower res in highlights)

e) & one more: Are you not curious as to whether the CCD sensor in your astro-camera has lower noise than those used in DSLR's? (Not lower noise when cooled, compared to a DSLR at ambient, but lower noise compared to a DSLR - both at ambient?)

Not only did I not receive a single answer to any of my questions, this was his disappointing response:

Dear Mr Lord,

I got your email today. Initially, I wrote a lengthy reply addressing several of your points. I then wrote another version which was somewhat more terse. Finally, I wondered why I should bother at all.

.................... I stand by the view that pixel-pixel variation, read noise, bias, thermal noise, spectral response, exposure duration and temperature are key. The DxO info supplies only a subset thereof - and does roll pixel-pixel into "noise".

So, let me explain it to you. Feel free to think me an idiot. Feel free to write your own articles and get them published in magazines, journals and books. Feel free to do pretty much anything you like. Frankly, I will not give a damn.

Everything I was told about you was true. What a rich pageant life is.

Don't bother responding. Your email address is now filed under - "Send to Recycle bin".

Farewell.

Grant Privett

Clearly someone who has an overblown sense of his own importance, feeling under pressure to defend it. The last question, (e) is the significant one. What amateur astronomer purporting to be curious about the world about him or her, and curious about the technology they use to investigate their area of interest, would intentionally decline the opportunity to find out whether the ccd sensor they use is less or more noisy than a DSLR cmos sensor? Well evidently Grant Privett does. What can you say about an amateur astronomer who refuses to rise to the challenge of making such a comparison? Amateur astronomy is a scientific hobby. Its very essence ought to be imbued with scientific curiousity. What we have here is a crystal clear example of the purbllnd, the intentionally uncurious, more concerned with defending their status, that getting at the truth.

There is an old addage, "There are none so blind as those who will not see" et tu Grant Privett.

Chris Lord

This page was created by SimpleText2Html 1.0.2 on 20-May-2010

Top of Page
Return to Home Page