PaulRivers
Veteran Member
Joe0Bloggs thanks for posting those shots - it's interesting.
-
Welcome to the new forums! Please read this first. For known issues we are working to resolve, click here.
shotting at lower resolution
- Thread starter doremon313
- Start date
I'm not 'ignoring' anything as such - it is not my intention to make full and complete analysis of every aspect.
This is primarily a discussion regarding 'ISO' with respect to 'signal levels / sensitivity' - noise is a 'secondary distraction', a side issue, to the main point of my discussion/contention.
I don't really care about the noise level (for this discussion) - I'm talking about 'signal level' with respect to 'ISO'.
That is quite wrong - if you add noise to the data (by randomly filling LSB zeroes) - you simply add noise - nothing is improved by adding random noise to the data post capture.
You can 'sort of' reduce quantisation effects by adding random noise to the actual signal before quantisation In signal processing terms this is often called 'dithering' - however, it is normally only effective when data is ' averaged over time ', for example in 'image stacking' (e.g. astro-photography), and not quite so effective in a single capture. More often than not, there will be no need to 'add noise' because there will usually be more than enough noise present anyway.
Joe0Bloggs
Veteran Member
Signal means nothing without noise. If you double the signal but also double the noise, you have done nothing for the signal to noise ratio.
ISO is not defined by 'signal level', but by the light level that corresponds to middle grey in the output image. The ISO standard does not say anything about what is done to achieve this end result. Thus software ISO boost and binning are equally valid approaches to 'high ISO' and the ONLY question is whether one has better signal-to-noise ratio than the other.
The binning approach has better signal to noise ratio not because it 'uses more bits of the ADC' but because you pay the noise penalty from the amplifier, A/D converter, etc. just once for four pixels combined rather than once for each pixel.
As I pointed out, the least significant bit in the highest 'physical' ISO setting is already way below the noise floor--if the amplifier had extra higher settings, it would simply lift the noise floor higher up the ADC bits and the extra bits showing up under the LSB of the original highest 'physical' ISO setting would be pure noise.
I didn't say that filling in the zeroes with noise would 'improve' anything, I just said it would exactly replicate the effects of a higher physical gain.
The image with zeroes loaded in Photoshop would have gaps in its histogram (not accounting for the effects of colour mapping) while the image would noise filled in wouldn't--but neither image would look better than the other because they both have the same signal to noise ratio and the posterization occuring with the first image occurs under the noise floor.
I know that full well - there is really no need for you and Ron Parr to keep endlessly repeating it.
It really does not matter (to my original point) if the noise doubles with the signal - how many more times must I say this?
The original issue/contention was whether or not 'binning pixels' equates to increasing 'sensitivity' and/or increasing 'ISO' - it was not about 'SNR'.
As far as the photographer, and the camera's internals, are concerned, 'binning' increases signal levels - which certainly equates to increased ISO, and debatably could be described as a form of increased sensitivity (greater signal out for same energy input).
As I've said many many times - noise is a secondary consideration.
It wouldn't even matter to my arguement if the SNR got worse - of course it would bother the photographer - but it's not key to my original point.
Thankyou - that pretty much validates my original point.
No - had the original point of the discussion been SNR then I would agree - but SNR WAS NOT the original point.
Again - my original point WAS NOT 'SNR'.
The binning approach is actually 'good/better' for several reasons - averaging of random noise, single read noise, and better utilisation of the A/D range.
It looks like I initialy misunderstood the intent of your illustration.
However, I still wouldn't agree with the point your are describing...
You are 'effectively' saying that there is no point in amplifying any signal once the noise level reaches/exceeds 1 bit, the LSB (which of course is ludicrous), because the noise is just pulled up with the signal.
If that were actually true, then few, if any cameras would ever need variable gain amplifiers.
It is not likely to be true that they would always look the same at all.
4 bits of noise level in 10 bits of A/D range will not look anything like 0 bits of noise in 6 bits of A/D range - that's an extreme example, but it's to make my point.
The fact is, that there generaly needs to be at least around about 10 bits of A/D resolution wherever possible for decent image quality (avoidance of posterisation like effects) almost regardless of the level of noise present.
Believe me, a 'smoothly resolved noisey image' will be much, much, better than a 'posterised noisy image'.
Joe0Bloggs
Veteran Member
Here's an ISO 3200 image taken with my Canon S90, processed in Camera RAW with noise reduction and sharpening turned off. One half was then untouched, the other half had its output level set to 0-15 and then the input level set to 0-15, thus effectively converting it into a 4-bit image. Can you see which is which without loading the image in Photoshop?
I don't dispute that binning pixels is a way to increase sensitivity and ISO. It's just that multiplying the ADC values after the fact would have the same effect of increasing sensitivity and ISO if the former method didn't yield a better signal to noise ratio in the end.
Absolute signal level has no meaning on the recording end, after all--I'd agree with you if we were talking about playback, where all the SNR in the world would be meaningless to the listening experience if your speaker had a max output of 1dB...
I don't dispute that binning pixels is a way to increase sensitivity and ISO. It's just that multiplying the ADC values after the fact would have the same effect of increasing sensitivity and ISO if the former method didn't yield a better signal to noise ratio in the end.
Absolute signal level has no meaning on the recording end, after all--I'd agree with you if we were talking about playback, where all the SNR in the world would be meaningless to the listening experience if your speaker had a max output of 1dB...
Joe0Bloggs
Veteran Member
Here for comparison is a similarly processed set of images of the same shot taken at ISO 80:
No trouble seeing which is the 4-bit image here.
No trouble seeing which is the 4-bit image here.
Joe0Bloggs
Veteran Member
no text
Joe0Bloggs
Veteran Member
I forgot that RAW data does not have a gamma curve applied, thus in the gamma-adjusted image the digitized levels are not evenly distributed along the histogram. I simulated this by applying a gamma of 0.67 before quantization and 1.5 afterwards.
Updated images:
ISO 3200:
ISO 80:
Updated images:
ISO 3200:
ISO 80:
I have to admit, your ISO3200 sample(s) have demonstrated your claim very well indeed.
I find that it is 'just about possible' to tell the difference, to see which one is which - but it is certainly quite difficult to see without very close scrutiny.
I give you full credit - but I would like to just point out one or two modest caveats...
One - the actual noise level of the ISO3200 RAW image data is extremely high indeed - I found as much as 140+ noise values in places in the blue channe - so it's not surprising that the noise level is several bit levels higher than the LSB level, even with only 16 levels.
Two - if we were to knock just 1 more bit out of the 'simulated lowered A/D quantisation' (down to 8 levels ) then the difference between the images becomes very obvious - which illustrates that some minimum relationship between quantisation step size and noise levels, does need to be maintained at some point or other.
Thanks again for your work and input.
P.S. I hope to get back to that other 'blockiness' '& 1/9th res' part of the thread sometime later/most likely tomorow - I'm sure I've spotted some things 'right' & 'wrong' there, for 'debate'.
I find that it is 'just about possible' to tell the difference, to see which one is which - but it is certainly quite difficult to see without very close scrutiny.
I give you full credit - but I would like to just point out one or two modest caveats...
One - the actual noise level of the ISO3200 RAW image data is extremely high indeed - I found as much as 140+ noise values in places in the blue channe - so it's not surprising that the noise level is several bit levels higher than the LSB level, even with only 16 levels.
Two - if we were to knock just 1 more bit out of the 'simulated lowered A/D quantisation' (down to 8 levels ) then the difference between the images becomes very obvious - which illustrates that some minimum relationship between quantisation step size and noise levels, does need to be maintained at some point or other.
Thanks again for your work and input.
P.S. I hope to get back to that other 'blockiness' '& 1/9th res' part of the thread sometime later/most likely tomorow - I'm sure I've spotted some things 'right' & 'wrong' there, for 'debate'.
Keyboard shortcuts
- f
- Forum