70D Dynamic Range is actually great, despite what DXOMark says

Started Jan 21, 2014 | Discussions thread
TTMartin
TTMartin Veteran Member • Posts: 7,304
Re: 70D Dynamic Range is actually great, despite what DXOMark says

David Hull wrote:

aftab wrote:

David Hull wrote:

Just another Canon shooter wrote:

David Hull wrote:

So here is a question for you: As you say, the effective system noise (camera noise, if you will) is higher at ISO 100, so tell me how Canon would fix it. in other words, based on the DxO data and the Sensorgen information based on it, what part of the Canon architecture needs to be improved?

I called that read noise, not system noise, and they should improve the readout circuit. It is trivial and not worth discussing.

When you measure it, you do not know where it is coming from (but we have additional info indicating what is going on). What Sensorgen does it to model it as additive noise and a multiplicative one, and fit a curve. Then they report the additive noise as "read noise".

He isn't really modeling anything, he is taking the DxO information and presenting it in a different perspective, one which makes it a bit easier to see what is really going on in the camera. It is certainly read noise but it is not coming from the sensor, it is not trivial and it IS worth discussing for those who want to understand what is really going on in the camera. If you want to understand that DxO DR curve and why it looks the way it does, you have to take the time to understand the camera's electrical implementation, the noise line up etc.

I think one of the things that Canon grapples with is that they do have a very good sensor technology (and have had it longer than anyone else). Their chosen implementation is sub par to the competition in two areas which manifest primarily at low gain settings. However, at the other end (where probably the bulk of people use the camera) they are fine.

Some of us want to know the why behind the what, for those it probably IS worth discussing. My point was simple, the sensor is just part of the imaging system, and in the case of Canon, it does not appear to be the root cause of the low ISO deficiencies that everyone like to point out in these threads.

Emil Martinec in his 2008 article shows what you two are discussing about. Read noise coming from circuitry upstream of the ISO amplifier ( I think this is what you are calling sensor read noise) and read noise coming from circuitry downstream of ISO amplifier. Canon seems to have problem with the later part. Sony and other manufacturers (Panasonic, Toshiba etc) have addressed this with on chip column parallel ADC. From memory Canon has a patent for on chip column parallel ADC since 2007. My feeling is that they haven't implemented it yet because they want to refine the technique as much as possible, add newer innovations and wait for the time when it would be cost effective.

Some have postulated that they cannot implement it within their own fab. I suspect we will see something from Canon in the not too distant future. I doubt that they are going to like giving up this spec to Sony for long. Exactly what they do is the question. I think that what keeps them going is that their current architecture is good enough since more people are concerned about clean performance at the high end. If the press is to be believed though, the new Nikon may be applying pressure on that end now. With an ISO rating of 400k+, 25600 is probably pretty impressive.

It should be pointed out that all of this is just a theory.

If Canon's problem was truly that their 'off the shelf ADC' was too noisy so it was losing 2 bits to noise. You would think that for a camera like the 1D X Canon could simply use a 16 bit off the shelf ADC and still have 14 bits noise free.

In the Magic Lantern white paper on dual ISO they state: 'We’ll focus on the uncompressed 14-bit data found in Canon’s image buffers from the camera itself. This format is used by Canon code when saving still photos, and it is codenamed MEM1 in the firmware; some relevant strings are sdsMem1ToRawCompression and sdsMem1ToJpegDevelop.

. . .

Raw values range from some black level to some white level (usual values are 2048 and 15000) and they are linear. To convert them in stops (EV), subtract the black level and apply log2:

ev = log2(max(raw - black, 1)) (1)

where ev will be a value between 0 and almost 14. To convert from ev back to raw values:

raw = 2ev + black (2)

with some extra care to avoid going outside the [black...white] interval.

The raw image usually contains two black bars, at top and left side, which do not contain imagedata, but can be used to estimate the black level and to analyze the noise.'

Bolding added.

 TTMartin's gear list:TTMartin's gear list
Canon PowerShot S100 (2000) Canon PowerShot SD1000 Canon PowerShot G1 X Canon EOS 40D Canon EOS 400D +26 more
Post (hide subjects) Posted by
(unknown member)
Keyboard shortcuts:
FForum PPrevious NNext WNext unread UUpvote SSubscribe RReply QQuote BBookmark MMy threads
Color scheme? Blue / Yellow