Well, sort of. When you read the sensor pixel you get a value in the range 0 to 2^n-1 where n is the maximum bit depth. If more photons are received the value stays at 2^n-1 (just as you can't raise the water level in a bucket once the bucket is full).

When you amplify the reading part of this range is mapped onto 0 ... 2^n-1 by a linear transformation. x2 if twice the native ISO is set, x4 if 4 times etc. If the resulting value exceeds the limit the maximum value is set ... so there is more clipping at high ISO than at native ISO.

Finally the value is transformed by the gamma function if shadow fill in is set - this gives a higher multiple at low values. If the image is then converted to jpeg, after compression the resulting values are linearly mapped to the range 0 ... 255 since jpeg understands only 8 bit depth.

And photons are always detected or not detected - you can't detect half a photon. The number of photons you detect at a pixel, if you repeat the process many times and ignoring thermal noise etc, will be a Poisson distribution centred on the average ... the deviation from the average in any particular exposure is what causes quantum noise, which therefor increases as the ISO is turned up (because the deviation is multiplied along with the raw value).

This is a considerable simplification because of offsets, thermal noise etc. plus variation in sensitivity between individual pixels. For a full analysis, look up AIP (as mentioned above).

 

This raises a interesting question, why don't camera maker use more bits ie 16 bit instead of 14?

Would that not improve things?

It's one of the difference between Nikon and Sony which are using the same sensors.

I know it would increase the data per shot but still if it gave better noise handling.

 

In terms of dynamic range, yes. But there is a cost, related to the accuracy of the readout logic, especially when there is a requirement to read out the sensor quickly because of frame rate / live view / video considerations. Dedicated astro cameras using similar sensors (and usually with a lower pixel count) take their time reading the data out, readout times of up to 30 seconds are common (and not an issue when the exposure times are measured in minutes anyway); the better ones achieve resolutions in excess of 15 bits from what is still basically 16 bit hardware. Going to 24 or 32 bit pixel depth would increase the amount of support circuitry considerably, increasing the manufacturing expense of the sensor and reducing the area available for detecting photons, and in any case would not give much improvement in a general purpose camera with a requirement for rapid sensor readout.

 

Because we can determine the sensor noise floor , surely there is a way of masking it (noise) just like Ray Dolby did within the world of audio with the Dolby noise reduction system. Obviously sight is different to sound, but perhaps there is a way of recognising actual image data versus system noise and boosting it to mask the noise?

Just a thought!!

Ian

 

That's effectively what Prof. Newman's article was about - increasing exposure of low light scenes to improve the signal to noise ratio. This is in addition to the camera's (or post processing) noise reduction algorithms, which try to remove noise without losing too much detail.

 

It's a lot more than a thought, there are some quite sophisticated ways of doing this, but there are two major issues. One is the amount of computational power needed, the other - rather more important for general photography - is that noise reduction has to be done late in the processing chain in order to be effective.

The distinction between "noise" and "data" is statistical in nature and therefore you can't be sure that anything you see in the finished image isn't a processing artifact or that significant detail isn't removed from the image. Some early Nikon DSLRs had a primitive noise reduction system which worked fairly well in many circumstances but had the unfortunate effect of deleting most of the faint stars when used for astrophotography - the dreaded "star eater" is gone now (or at least is turned off by default) but is still responsible for massive prejudice against Nikon amongst the astro photography community.

 

Thinking about this some more, it should be at least theoretically possible to automate the "expose to the right" system. Cameras with live view are reading the sensor at least 30 times a second (although possibly not at full resolution), and with fast enough signal processing, it should be possible to adjust the exposure until the brightest part of the scene just reaches the clipping point. The pixel values could then be scaled back to the nominal exposure after capture, and noise would be scaled back with them.

Since, as far as I know, the limit of Moore's law has not yet been reached, the cost of computer power is still falling, so although this might not be economically viable at present, it could be possible in the future.

 

That's pretty much what evaluative metering is doing.

Iyt is however necessary to leave some headroom, sharpening (in particular) tends to increase some of the pixel values by an amount which can be very significant.

 

Wouldn't it be good if the manufacturers would comment about our thoughts.
Surely they have already investigated stuff like this?..

 

Maybe the next time The Mighty Tharg has a chat with manufacturers?

Perhaps I should patent the idea first...

 

I would claim that it was my intellectual property, but I am not intellectual in the slightest !.