High ISO, analog or digital multiplication ?

a recent discussion on getting the best noise images from a pentax DSLR has spun off into a tangent of high iso ana/digi discussion.

let's continue that discussion here for those interested.

tnx

jordan

I can see Gordon was totally right, I really did make a mess in the last thread.

So I'll take responsibility and clarify my position right from the start.

In the previous thread I tried to make two points:
1) It makes a difference using analog vs digital amplification.
2) The difference isn't that significant.

It appears that I made point 1 more strongly than point 2 even though I mentioned near every post that it has limitations and questioned its significance.

That said, I'd also be happy to participate but its more interesting for EE than photographers.

Real Engineer wrote:

I can see Gordon was totally right, I really did make a mess in the last thread.

So I'll take responsibility and clarify my position right from the start.

In the previous thread I tried to make two points:
1) It makes a difference using analog vs digital amplification.
2) The difference isn't that significant.

It appears that I made point 1 more strongly than point 2 even though I mentioned near every post that it has limitations and questioned its significance.

That said, I'd also be happy to participate but its more interesting for EE than photographers.

how about for EE's who are also photographers ?

Cool.

IMHO, the main benefit of using an analog amplifier vs a digital one is when stacking multiple images. While any single image has very small benefits as the data increase is either unobservable or near unobservable, when the images are stacked this small difference is pulled out of the noise floor.

One such application would be astronomical photography.

This field of photography pushes the limits of the instruments simply because to get a usable image it nearly always has to be pulled out of the background noise.

If you've ever tried to take a single exposure of any astronomical object you'll quickly find that single images even with long exposures have their limits.

Most astro photographers now use a combination of long exposures and stacking to get the image quality that gives that "wow" factor.

If the image was stacked using only digitally amplified images there is no way that the resulting stacked image can have any benefits of the increased quantization levels that would be present if the images were stacked by using analog amplified images simply because the data was never captured.

How much difference that will be is going to be application specific and variable depending on the amount of images stacked, the subject matter and the characteristics of the camera and lenses.

glanglois wrote:

How about non-EEs who (in a limited fashion) understand the analog vs. digital world and are eager to learn more about both photography and signal processing??

In other words, we have a lot of different people reading this forum and almost all are here to learn.

I realised after my initial post I may have come across as elitist (which I'm not) so I've steered the discussion to an application where it is more photography and signal processing related.

In a somewhat related issue regarding high ISO, I've been thinking about the fact that Pentax does NR on ISO3200 and ISO6400.

Considering the fact that both are internally digitally generated from the analog amplified ISO1600 I no longer consider it an issue that Pentax does NR on these ISO settings. The reason for this is that both can be generated without NR simply by under exposing the ISO1600 image.

Basically they have not reduced functionality of the camera as I initially believed. What they did was provide an easy method for the average photographer to have higher ISO's with a little NR thrown in to make those images look better.

For the more demanding photographer it is still possible to get those ISO speeds without NR simply by limiting their shots to the analog amplified ISO settings and PP.

Its just a question of convenience to one group or another and Pentax choose to make it more convenient for the less demanding photographer.

It would be an interesting to see if the images from Nikon and Canon at the higher ISO's are analog or digital. If they are analog then what we're comparing are images from digitally shifted data to ones that are analog amplified at settings of ISO3200 and 6400. IMHO this should be pointed out in a review when doing the comparisons and DXO should also state which ISO's are digital and which are analog as part of their testing procedure.

I'll just qualify the above that I'm commenting from a theoretical perspective which is currently at about the limits of my my knowledge on the subject.

I haven't tested the theory to see what the limitations nor have I confirmed what I suspect would happen regarding the stacking differences between analog and digitally amplified images.

If someone has done this or see's a flaw in my logic then please take the time to point it out.

opiecat wrote:

a recent discussion on getting the best noise images from a pentax DSLR has spun off into a tangent of high iso ana/digi discussion.

let's continue that discussion here for those interested.

Hello Jordan, I think this is a good idea as the concepts of "real" vs. "fake" ISO sensitivity gain as used to produce the effect of high ISO's using digital sensors by means of analog amplification vs. digital multiplication are often disputed and yet have no real significance in considering the concepts of Exposure To The Right (ETTR) which was the subject of that thread. To that end, rather than making a link back to what I said in that thread, I here copy excerpts from the relevant parts of what I said as a starter, with some expansion on those ideas:

First, the evidence that digital gain has been carried out is that there will be underpopulated values in a raw histogram. For instance, the very simplest scaling of the normal "full stop" ISO's of 100, 200, 400, etc. (not that I would recommend that for all cameras for reasons I will explain later) could be done by multiplying by two using simple digital binary left shift operations, which would successively produce gaps in a raw histogram for the odd values, then populated values for only every fourth value starting at 0, then populated values every eighth value starting at zero, etc. However, Pentax cameras starting with the K20D also do other digital scaling in order to "idealize" the range so that the maximum "white point" is adjusted away from its "native" value to to the full scale 12-bit range of 4095. Thus, for the same adjusted exposure of all ISO's between 100 and 1600 for the same Imaging-Resource K-7 Still Life raw scene as can be downloaded from: http://www.imaging-resource.com/PRODS/K7/K7THMB.HTM , the raw histogram as produced by my adapted Java DNG Analyzer program available from here: http://cid-610bb9d47b109171.skydrive.live.com/browse.aspx/Distributed%20Programs looks like the following for ISO 100:

and like the following for ISO 1600:

I use my Java DNG Analyzer for this rather than PhotoBola Rawnalyze because the resolution is higher, yet generates GIF's rather than JPEG's the file sizes are not large.

Note that the only differences between these raw histograms other than slight differences due to the absolute differences in exposure are the spacing of the gaps that has been caused by the digital scaling Pentax has done: lower ISO's have not had to be scaled as much and higher ISO's by more due to Pentax also having compensated for the varying black offsets in the originating raw histograms before output to the raw file format. In a way, Pentax have already used digital amplification but it is here only about a factor of 10% for ISO 1600 and less for lower ISO's, so no really as extreme as what is being debated.

To repeat what I posted on the previous thread as proof that the Pentax K-7 does do more digital scaling for the ISO 3200 Still Life image as there are underpopulated raw values roughly every other value, I posted the following, which also shows the effect of the raw Noise Reduction (NR) carried out for ISO 3200 and higher as sub patterns in the histogram:

and for completenesses sake, I will also post the ISO 6400 version showing that the underpopulated gaps in the raw histogram is again twice as extreme although the sub patterns due to raw NR are about the same, as follows:

Regards, GordonBGood

This will be continued in successive parts with analysis of how little damage digital amplification really does, and it's limitations in those situations where it isn't so good.

GordonBGood wrote:

This will be continued in successive parts with analysis of how little damage digital amplification really does, and it's limitations in those situations where it isn't so good.

To continue with an analysis of the effects of artificially increasing quantization error by reduced encoding levels as follows:

For the following, I will use Raw Therapee - RT - (www.rawtherapee.com) as a raw converter, not because I particularily recommend it, but for verifiability as it can be downloaded for free and because, now being open source, a competent programmer can verify that there is no special treatment as to demosiacing, colour tone curves, etc. that is specific to any camera model, making it "camera model neutral". I will use completely neutral processing with absolutely no Noise Reduction (NR) or sharpening and will set all development sliders to zeros except as noted.

As proof whether high ISO gains are achieved by analog or digital means makes no perceptible difference for most cameras in images produced from such gains because the base level of noise is much higher than the gaps in the codes, I offer the following method:

As we have no cameras that do this gain by both digital and analog means, it is necessary to simulate a digital gain from an analog gain in order to prove this. For instance, given from Part 1 above that the K-7 produces ISO 1600 images using analog gain, and where I have measured the lowest level of noise in the blacks at about a standard deviation of 22 to 29 12-bit levels (depending on the camera sample), one could take any K-7 ISO 1600 image and increase the quantization error by rounding down all of the raw odd values leaving a zero population of odd values as if the raw data had been created from a "real" ISO 800 image that had been digitally gained by an exact factor of two. Further, one could round down in groups of 4 codes (binary anding with zero's in the least two significant bit positions) to leave gaps of three codes in the raw histogram as if it had been digitally scaled by a factor of four from an ISO 400 image. Further, we could use the same method to eliminate all but every eighth value and all but every sixteenth value as if we had produced the ISO 1600 image by digital multiplication from an analog produced ISO 200 and ISO 100 image, respectively. I have done this to sample images using a progam and written the results out to a DNG file. Upon normal identical raw conversion of each of these five images, I don't think that anyone will see the difference in real image quality due to any loss in quantization, as these effective code gaps are filled by processing in the later raw workflow due to the very lowest level of noise in the ISO 1600 image, the black read noise, being much higher than even these large gaps in the codes.

For your comparison here are the results:

The original ISO 1600 image (original size in my gallery):

The "quantized" ISO 1600 image as if it had been produced from a ISO 100 image with black noise proportional at ISO 100 to what it is at ISO 1600, which is the same other than being a tiny bit darker due to always rounding down (original size in my gallery:

As I think you can't see any difference, this proves that quantization does not produce visible quantization step errors or "banding", even in the very dark tones when boosted, as long as the high levels of minimum noise levels are bigger than the stepping caused by the quantization.

To further push this concept, I have used RT to push the quantization level starved raw image that has code gaps of 15 code levels for every one that exists by an additional raw processing EV boost of 4 stops for an equivalent of ISO 25,600 and view it with RT at 100% zoom for the very darkest area of the above image. Essentially, because I eliminated 4 Least significant bits from the raw file and then boosted by another factor of 16, I have increased the quantization level so there are equivalent gaps in the raw histogram of 255 codes for every one that exists or only 4 bits of usable data, and also multiplied this level up by the factor of 16 to better see any stepping of quantization/"banding". I have shown the darkest area I could find in the image that has a smooth gradient as since the development Tone Response Curve (TRC, as per "gamma" curve) is steepest in the dark tones, it will most amplify any steps in the linear raw file as larger steps in the developed image. This worst case image that pushes the use of quantization to the maximum looks as follows:

I don't think you can see any gradient stepping buried in the noise.

Continued in Part 3, where we push this even further...

Regards, GordonBGood

Continued from Part 2...

The reason that producing high ISO gains by digital means is absolutely a non-issue for high ISO's from the K-7 is because the high base level of noise is much larger than the gaps in the raw histogram and tends to fill in the missing gaps in further raw procesing so that we don't see bands in the smoothness of the tone curves. According to Dr. Emil Martinec (ejmartin in these forums), we probably don't need to be concerned with gaps in the raw histogram as long as the base noise is at least twice to three times bigger than the space between the quantization levels, and that looks to be true. Thus, for a very low noise camera such as the K-x where the noise at ISO 1600 is only about 6 12-bit levels, the more extreme limits of this test would likely start to slightly degrade the image quality, especially if we boosted the EV in post processing so we could peer into the darkest levels where the gaps are most meaningful. However, those very low levels of noise are only present in the very darkest image tones and by the time we get up to four or five stops below clipping, again they are a non-issue.

Here is the same experiment done on a K-x ISO 1600 image with quantiztion as if it were a boosted ISO 100 image to try to reveal the "banding" quantization in the darkest area boosted another 4 stops using RT:

You might just be able to be able to see the stepping in the darkest image tones as seen here, but again such stepping is fairly effectively blended together by the raw processing of the noise levels which have masked the stepping and it is hardly a problem.

As you can see, reducing quantization levels is not a problem as long as the quantization steps are not smaller than the amounts of noise in the images. In the next part, I will explain the real reason that analog amplification is necessary for better high ISO images for some cameras.

Regards, GordonBGood

Real Engineer wrote:

I'll just qualify the above that I'm commenting from a theoretical perspective which is currently at about the limits of my my knowledge on the subject.

I haven't tested the theory to see what the limitations nor have I confirmed what I suspect would happen regarding the stacking differences between analog and digitally amplified images.

If someone has done this or see's a flaw in my logic then please take the time to point it out.

I'm afraid your logic is indeed flawed with respect to having to treat stacked images differently as to requiring analog or digital multiplication for high ISO images. As I have proved in my posts here, the only effect of using digital multiplication instead of analog multiplication is to reduce the quantization levels, and that this reduction of quantization levels does not matter when the image noise is at least as high as the gaps in the codes. The effect of stacking images by averaging their values is to reduce their noise such that stacking say 16 images will reduce the per pixel noise by a factor of four. However, another effect of stacking a quantization-level-reduced image is to increase the effective bit depth. For instance, think about a level that should have been quantized as level '7' but the quantization had been decreased by two bits so that the level would be represented by either a '4' on one side or a '8' on the other. That isn't significant if the level of noise is higher than 4, as out of those 16 stacked images, it is probabe that there will be about 12 images with that particular photosite having a level of 8 or higher and about 4 images with that particular photosite having a value of 4 or lower, with the effective average actually tending to recover the lost two bits and making them more meaningful.

Thus, the effect of averaging a stack of identical images other than the noise will be to tend to fill the gaps in the raw histogram as if they hadn't been created in the first place. As always, this only works when the image noise is sufficiently larger than the gaps in the quantization levels, but that is always a requirement for effective use of digital multiplication.

Regards, GordonBGood

GordonBGood wrote:

Real Engineer wrote:

I'll just qualify the above that I'm commenting from a theoretical perspective which is currently at about the limits of my my knowledge on the subject.

I haven't tested the theory to see what the limitations nor have I confirmed what I suspect would happen regarding the stacking differences between analog and digitally amplified images.

If someone has done this or see's a flaw in my logic then please take the time to point it out.

I'm afraid your logic is indeed flawed with respect to having to treat stacked images differently as to requiring analog or digital multiplication for high ISO images. As I have proved in my posts here, the only effect of using digital multiplication instead of analog multiplication is to reduce the quantization levels, and that this reduction of quantization levels does not matter when the image noise is at least as high as the gaps in the codes.

Makes perfect sense.
I also see where my logic went wrong too and should have known better.

Its one of the cases where a little system noise actually increases the effective resolution of a ADC system and considering I come from an analog and digital hardware background where it is also quite well understood and documented is well, rather embarrassing.

For completeness and to use my mistake to help others from making the same, I'll explain what went wrong with my logic.

Quite simply I made a classical error.
The logic was like this:
Stacking reduces noise.
The quantization gap stay the same across all images using stacking.

Hence the use of analog amplification will satisfy the noise gap ratio for increased detail.
Obviously flawed.

Thanks mate.