theSuede
Offline
• • • •
Upload & Sell: Off
|
I'm not Tony, but some of the questions are quite easy... (well, answerable - almost the same thing!)
- You gain just as much in the bright end, but when the signal is strong the relative difference that the improvement makes is very small. The effect is weak in the bright end, and strong at black-level - compared to the signal.
- No, raw-files are always (roughly) linear, no tonemapping or gamma has been added to the values. The value "100" over blacklevel will always mean a brightness twice that of the value "50" in a raw (except for the compressed raws...)
You have to grasp the concept of "noise-adding" or "noise-averaging" to know why certain things have the effect that it does. There are several "stages" of noise added to the signal while it is being transformed starting from "X" amount of light hitting the sensor surface resulting into a "N" value in the raw-file.
In short, 14-bit makes more difference in high ISO's because it mostly affects the lower effect read-outs from the sensor - and those lower effect readouts are amplified further up into the picture exposure if you amplify the signal more, i.e. choose a higher ISO. a read of say 20 electrons means one brightness level on base ISO, but twice the brightness on next ISO and so on. If base ISO is "100", the ISO1600 means that you only use the lowest 1/16 part of the available light gathering capability of the sensor, the equal brightness values comes from amplification, not measurement. This of course means that you amplify readnoise 16 times also...
Roughly, the noise sources are:
-1. "ShotNoise SN" - The incoming light is not "perfect" - it contains an uncertainty part that is equal to the square root of the measured amount. It the sensor measures 4 electrons, error margin (Shot Noise) is 2 electrons. Measure 9, SN is 3. Measure 100, SN is 10.
-2. "ReadNoise RN" - The number of charges that you collect has to be "moved" from the photosite to be converted to a voltage that an analog-digital converter can convert into a digital number. This "moving" includes a certain error-margin - which is the same as noise.
-3a. "AmplificationRatio AR" - Signal amplification multiplies the sum of the above. Both the signal and the noise is amplified equally.
-3b. "AmplificationNoise AN" - Signal amplification adds a bit of error margin in itself while amplifying the signal. This noise is added after amplification and is not included IN the amplification.
-4. "QuantizationNoise QN" - this is an effect of converting an analog signal into discrete numbers. The short explanation is that if the value measured is "55.5", the ADC has to choose either real number "55" of "56" to output, giving an error of ~0.5 to the value. This effect is very low in the overall noise adding (12 bits are ~4095 levels to choose from..).
With that done, it's time to start thinking about what changing different parts of the chain will change in the end result.
-1. The only way to improve shot noise is to increase the number photons converted. You must improve the Quantum Efficiency of the sensor (how good it is at converting light energy to electric energy), or increase the exposure (and this is not always possible). A normal QE value for a modern camera like the D300 is about 35-40, ie around 35% of the light hitting the sensor gets converted to electric charges.
improvement here will lower noise in ALL exposure levels and ISO's by the square root of the improvement.
-2. Read noise is fairly "static" - it does not change with amount of light converted. This means that it mostly affect shadows, highlights will show no change as the noise AMOUNT is bigger in the highlights (but less visible as the signal level is higher!). This is added to the shot noise, the sum equal to √(SN^2 + RN^2)
improvement here will only affect exposure levels where the read noise is a significant part of the sum, ie the collected light amount is very low. This is the extreme shadows, at least 10Ev down from saturation at the lowest amplification - i.e. at base ISO.
-3a. Does nothing more than multiply the signal from "-2.", signal and noise both equally.
improvement here is only possible by choosing a lower amplification (which will lower ISO), which means you have to get more light. This is not always possible... :-)
-3b. Amplification noise is added to the signal in the same way as read noise, by adding square roots of the active parts - but this is done AFTER amplification, which makes the effect equally strong in the shadows no matter what ISO you choose.
improvement here will affect the shadows on all ISO's, but it's not a very strong noise-source.
-4. This effect is in almost all applications lower than the minimum noise of the signal sum, so it's not very important as long as you're at 12-bit plus (12Ev SNR is about the limit of the best sensors today).
improvement here will ONLY make a difference on base ISO on the very best of sensors, in the values very close to zero (almost black)
Now what the "14-bit" readout does is to either do the readout slower (lowers RN and amplification noise - higher data rates always means more noise when you do signal processing. You get a more accurate readout.) - this is how the D3 and the D700 does it.
The D300/D3x and other SonyExmor based cameras seem to do a "multiple-read" - which they have to as the output from the sensors are limited so 12 bits by the AD converters on the sensor! - it takes the stored charges on the photosites and measure them several times. This will make the result more accurate by averaging the readouts and thereby lowering the effect of the RN and AN/QN by the square root of the number of sequential reads. 2 reads > 1.4 times less electronic noise "contamination", 4 reads > 2 times less electronic noise "contamination" (AND if you add 4 12-bit readouts you get - tada! - 14 bits).
Kinda long, but I was bored.
Edited on Sep 05, 2009 at 10:33 PM · View previous versions
|
