High ISO vs Underexposure. What is best for noise?

[Ray]

I understand that lower read noise at high ISO in DSLRs is due to analog preamplification prior to A/D conversion. In CMOS sensors for DSLRs I believe every photosite has its own preamplifier. I doubt whether there would be room for an analog preamp on the tiny photosites of P&S cameras. I would therefore not expect the same degree of read noise reduction at higher ISOs on P&S cameras.

I remember testing these issues with my 60D and was surprised to find that an ISO 100 shot underexposed by 2 stops had hardly more noise that an ISO 400 shot with the same exposure.

However, since the D60 Canon have reduced high ISO noise very significantly in subsequent models, yet noise and DR at base ISO has improved only marginally.

Whilst wandering around a temple today in the late afternoon, I thought I might try a few high ISO shots in poorly lit areas for stacking experiments. I wandered around to the back of an old building, deep in shadow as the sun sank below the clouds, and shot off 7 frames in continuous mode at ISO 1600, f8 and 1/20th sec. with my 5D, hand held.

I've just converted the images, checked the darkest parts for noise, and couldn't find any. I mean, how frustrating   . Perhaps I should have used ISO 3200 at f11.

Below is the ACR window displaying the conversion settings, followed by a very much lightened crop of the top right corner where I expected to see noise.

Needless to say, I'm very impressed with the low noise characteristics of my 5D. But I'd make one point; this image was correctly exposed to the right. The small areas of sky poking through the foliage are definitely blown, as indicated in the small spikes on the far right of the histogram. But the fact they are blown is inconsequential in this image.

Default sharpening and luminance smoothing of 25 was applied in ACR.

[attachment=3927:attachment]  [attachment=3928:attachment]

I should add that I'm on an uncalibrated laptop. I think all my recent images would therefore probably appear a few shades lighter when viewed on a calibrated monitor (according to my printer results).

[Guillermo Luijk]

I have done more tests, ranging all possible ISO values my camera (350D) has.



Shot at T=1/4, f/13, 50mm


ISO PROGRESSION
From left to right, lighter to darker areas.
Crops at 50% done using nearest neighbour to preserve SNR.



ISO100 vs ISO1600
Crops at 100%:




Improvement is so clear that noise manages to alter the general tone (colour) in the darkest areas.

The improvement is higher the darker the areas are. So in the darkest shadows the gap from ISO100 to ISO200 is clearly larger than from any other couple of ISO values. This is consisten with the theory that improvement is less noticeable once we reached very high ISOs.

[bjanes]

I have done more tests, ranging all possible ISO values my camera (350D) has.

The improvement is higher the darker the areas are. So in the darkest shadows the gap from ISO100 to ISO200 is clearly larger than from any other couple of ISO values. This is consisten with the theory that improvement is less noticeable once we reached very high ISOs.
[a href=\"index.php?act=findpost&pid=154487\"][{POST_SNAPBACK}][/a]

Guillermo,

Nicely illustrated tests. According to Roger Clark's unity gain theory, above ISO 1100 (the unity gain of the 350D), there is no reason to increase the ISO any further. You can achieve any further boost that is needed in the raw converter. This will give you greater headroom and improve dynamic range. He adds that most astronomy enthusiasts do not bother to go beyond ISO 1600 with Canon cameras.

On a camera like the Nikon D200 (unity gain at ISO 800), the results would be even more marked.

With a medium format digital back using the KAF 39000 chip with a 16 bit ADC and a full well capacity of 60K electrons, theoretically one could stay at base ISO and make adjustments only with the raw converter, but I have not seen any tests that confirm this.

Bill

[Panopeeper]

Guillermo,

I have downloaded the images from the first post, with ISO 100, 200 and 400 and analyzed them in raw format.

The ISO 100 raw image is less noisy than the ISO 400.

The stress is on raw. The noise is the consequence of the post processing, not of the original image.

A closer look at the raw images follows (not the ISO 200), but first some notes:

1. the black current level calculated from ACR based on the masked pixels is between 255 and 258 for both images. Everything under the respective black level is considered null (actually, ACR calculates one value for each row).

2. I selected a 100x100 pixel area from the top right edge of the monitor bezel. Many of these pixels, like in the other dark, noisy areas, are very close or even under this limit. However, the values with ISO 400 are in average higher, above the black level; this is very important.

One can't make very much out of nothing.

3. The following images are screen captures. The color is not the result of de-mosaicing but averaging of the pixels of a filter pattern (2x2 pixels).

4. In order to compare the raw images as "native" as possible, I have not applied any white balancing. No sharpening or any other hidden post processing has been applied.

4. Likewise, the exposure of the ISO 100 data is not adjusted.  Instead, the source range of the mapping was limited to 244-500, instead of 255-4095. Therefor the dark areas appear much lighter, without changing their values.

So, here is the raw image of the ISO 100 shot; the selected area is framed with an orange rectangle:



and here is the ISO 400 shot:



There are two row of numbers of the images, the first of them relates to the raw  values. They show the minimum, maximum and average pixel values within the selection. These show, that the ISO 100 image is more uniform (less noisy). For example the green of the ISO 100 shot are between 250 and 268, while the greens of the ISO 400 shot are between 246 and 291.

Well, if this is so, then where is the noise from?

It has two sources:

A. the black level.
Applying the black level correction "generates" many pixels with zero value. When the exposure correction is carried out, nonzero values get scaled, but it leaves the zero pixel values at zero.

Moving the mouse over the selected area in the ACR created images shows many 0 values as one color component. These are now noticable in the "new context", where the values of the non-zero pixels are quadrupled.

B. stretching of the scale. In order to achieve the same brightness with the ISO 100 image as with the ISO 400 one, a huge exposure correction is required. This stretches the pixel values in a four times larger space, which creates "holes" between the pixel values, and that appears as noise.

Here is a crop of the ISO 100 image with the statistics:



and the ISO 400 version:



The layered TIF of screenshots  (10 MB) contains these images without JPEG artifacts, plus screenshots from ACR and PS. The expanded histogram of PS can be used on a selection within the orange frame, which shows that the standard deviation of the ISO 100 shot (on that selection) is 5.2, while it is 8 in the ISO 400 image. The same on the ACR images shows the result of the minimum noise removal of ACR.

Note: the noise removal of ACR 4, 4.2 can not be completely turned off. This has been reduced in ACR 4.3, but there is still a small amount of noise removal, which is part of the de-mosaicing algorythm.

[Guillermo Luijk]

The ISO 100 raw image is less noisy than the ISO 400.

The stress is on raw. The noise is the consequence of the post processing, not of the original image.

Well it all depends on what we call post processing, and what we mean by "less noisy".

The ISO100 is less noisy, right, but has a worse signal to noise ratio, which is the parameter that really matters in order to be able to choose which image is preferred in terms of noise. To be fairly compared you must adjust RAW exposure so that bright levels on both images become the same. I don't consider this exposure correction as post processing, but anyway that discusion is all about semantics.

The question is the ISO100 has a lower signal level compared to its noise level than the ISO400 or the ISO800 so the ISO100 provides a less usable final image.

Yes, black levels on the 3 images was about the same: 256 levels. I always use DCRAW to get this information so as to make sure I have no noise reduction at all; something that as you say cannot be achieved using ACR.

BTW, what is this PhotoBOLA soft? mmm seems interesting.

Regards

[MichaelEzra]

PhotoBOla: I found it here: http://www.cryptobola.com/PhotoBola/Rawnalyze.htm

[Panopeeper]

The ISO100 is less noisy, right, but has a worse signal to noise ratio, which is the parameter that really matters in order to be able to choose which image is preferred in terms of noise

There is no question about this: if I have to choose between ISO 100 two stops underexposed and correct exposure with ISO 400, then I choose ISO 400, because it fills the gaps with useful data.

I don't consider this exposure correction as post processing, but anyway that discusion is all about semantics

I consider everything post-procesing, which is post, i.e. after the initial action. If I am recording raw data in-camera, then whatever I do with that data is post processing.

Sure it is about semantics, but it is about communications as well. Do you considere for example contrast adjustment in ACR post processing?

black levels on the 3 images was about the same: 256 levels

I think DCraw made it a bit too simple. It is the task of the raw processor to estimate the correction, based on the masked pixels. Unfortunately, those give very mixed, sometimes outright crazy information.

For example the masked pixels in your camera over the columns yielded 202-205 in average for every second column. However, the average for every other column is between 252 and 254. The individual values show huge variations, between 77 and 265 (only a cursery look).

ACR applies row-wise correction, calculated to be between 255 and 257 for the ISO 100 shot and between 255 and 258 for the ISO 400 shot. I have no idea, how (with what algorythm) ACR came to these results. The averages per row vary between 253 and over 3000 (yes, three thousand).

Perhaps the inherent noise reduction of ACR (which is part of the de-mosaicing) makes use of a more detailed analysis of the black noise data - although this non-optional noise reduction applies to such cameras as well, which do not deliver masked pixel values.

[Guillermo Luijk]

I think DCraw made it a bit too simple. It is the task of the raw processor to estimate the correction, based on the masked pixels. Unfortunately, those give very mixed, sometimes outright crazy information.

DCRAW estimates the correction by reading the level in hidden pixels, that's why it substracts a different amout of black levels for each individual RAW file. No idea how precise is the calculation of these levels, we should look at DCRAW's code or ask David Coffin.

When I have plotted undemosaiced 12-bit histograms of my RAW files I found no information below level 255:



If you substract black by 3000, you would eliminate most of the image! bearing in mind it reaches a maximum of 4095 (less in some other cameras).

[Panopeeper]

When I have plotted undemosaiced 12-bit histograms of my RAW files I found no information below level 255

There is plenty of data below 255. The ISO 100 image has 52494 red, 50245 green and 32533 blue pixels between 245 and 254. The ISO 400 image starts at 233 (only two pixels); there are 12634 red, 8930 green and 12656 blue pixels up to 254.

Following two images show the bpttom right corner. Note, that in this view not even the averaging of the pixels of the CFA arrays has been done; every display pixel represents the corresponding image pixel in the respective color, with mapped intensity.

(The pixel statistics in the image relate to the range up to 255, which I selected for the sake of displaying everything over 254 as black.)





The question is, how much of this data carries useful information, not pure noise. These images show the outline of the complete image, but that reflects only the fact, that the "undercutting" values occured in the very dark areas.

If you substract black by 3000, you would eliminate most of the image! bearing in mind it reaches a maximum of 4095 (less in some other cameras).

This is not so simple. In many raw images the clipping point depends on the ISO. The Canon 20D's clipping point is always 4095, like the 350D appears (I have not seen images of the 350D with all possible ISOs, but if ISO 100 clips at 4095, then most probably all ISO settings go up to 4095). However, for example with the Canon 40D, the clipping point is between 12700 and 16383, depending in the ISO.

[Guillermo Luijk]

The question is, how much of this data carries useful information, not pure noise. These images show the outline of the complete image, but that reflects only the fact, that the "undercutting" values occured in the very dark areas.

This is interesting Pano. And I don't think those data are only noise, at least not all of them. They can also be very low but useful levels, coming from pixels that got very little light excitation. In fact noise is surely some levels above in many pixels.
And if they are just noise appearing on super dark areas outlining them, that is information anyway, so why should we loose it?

Looking at the pure RAW histogram of the image you are referring to (my ISO100 test of the computer), I can see there is a soft end of the histogram near level 256, and as you say, there is quite a lot of non-zero information there.

I think the reason for substracting a black level being a power of two must be  simplification, I cannot find any other explanation. I will ask David Coffin but I don't think it was a problem to substract 246 (you can see I found the same value as you in my histogram for the beginning of the information , leaving aside a couple of G and B pixels with a 0 value):

Filled levels:  R: 3849 (5,9% of available), range [246..4095]


Zoom 1:1 12-bit histogram of RAW file showing clipped data due to 256 black substraction.


Definitively I am going to write to David Coffin to find out why he (and it seems ACR too) always substracts an integer power of two (I have developed some RAW files in which the black point substracted was 128, but 256 is more usual).
Luckily DCRAW allows you to develop setting the desired black point, in this case using: dcraw -k 246 would have saved those levels. But to calculate it in advance is time consuming.

PS: BTW if you like to investigate detailed histograms, and it seems you do, maybe you are interested in a little program I wrote called Histogrammar that plots histograms in the 0..65535 range. Also calculates logarithmic histograms for dynamic range calculations.

Download it from here:

Histogrammar Download (click DESCARGAR)

There is a tutorial although I guess you don't need it hehe: Histogrammar Tutorial

[Panopeeper]

I don't think those data are only noise, at least not all of them

Nor do I, and I am not convinced, that ACR (or even DPP) makes full use of every bit of information.

Definitively I am going to write to David Coffin to find out why he (and it seems ACR too) always substracts an integer power of two

This is not so with ACR. I uploaded three text files with the related values:

Masked pixels at the top, ISO 100

Masked pixels at the left, ISO 100

Vertical black level array, ISO 100

The latter shows the values generated by ACR (actually, by the DMG converter, but they are supposed to work the same way). ACR ignores the masked pixels at the top and creates a value for each row.

Note, that the 350D has masked pixels at the top and the left, but not at the bottom and the right.

I have developed some RAW files in which the black point substracted was 128, but 256 is more usual)

Not with the 350D, I guess. The average black level is around 128 with the 20D, 30D, 1DMarkII, 1DsMarkII, 5D, but it is around 256 with the 400XTi and 350D. I don't have any image from a 300D, nor from the 10D.

if you like to investigate detailed histograms, and it seems you do, maybe you are interested in a little program I wrote called Histogrammar that plots histograms in the 0..65535 range. Also calculates logarithmic histograms for dynamic range calculations

These look fancier, than the histograms created by Rawnalyze:



However, the space in a dialog window is a serious consideration when deciding the shape of a diagram. The histograms of Rawnalyze are 512 pixels wide, plus frame, etc. With some monitors this is nothing, but with some other monitors, it reaches the edge easily.

Another issue: I see a problem with the dynamic range without knowing the camera exactly. From where should the dynamic range be counted backwards? Most cameras' clipping points are not at the numerical limit of the respective bit depth, and it even depends on the ISO.

For example three new Canon cameras have now 14bit depth, but the maximum value is 16383 - or less than 13000, depending on the ISO. If I don't know the camera, I can't calculate in stops.

[Guillermo Luijk]

Not with the 350D, I guess. The average black level is around 128 with the 20D, 30D, 1DMarkII, 1DsMarkII, 5D, but it is around 256 with the 400XTi and 350D. I don't have any image from a 300D, nor from the 10D.

Yes, I think it was with the 350D, my only camera. Although it is true that many people has lent me RAW files, so actually you could be right.

However, the space in a dialog window is a serious consideration when deciding the shape of a diagram. The histograms of Rawnalyze are 512 pixels wide, plus frame, etc. With some monitors this is nothing, but with some other monitors, it reaches the edge easily.

Another issue: I see a problem with the dynamic range without knowing the camera exactly. From where should the dynamic range be counted backwards? Most cameras' clipping points are not at the numerical limit of the respective bit depth, and it even depends on the ISO.

I use a 768 (3*256 pixels) wide display, but you can zoom both axis from 1:1 up to 1:256, which is equivalent to Photoshop's 8-bit range histograms. Of course you can move to any point with a slide bar, it's easy and practical.

Regarding the dynamic range, the logarithmic mode allows to calculate the dynamic range of a scene visually by using a log scale in the X axis.
But this only makes sense when displaying linear images, like those from DCRAW. It always plots 16 f-stops histograms, which is the maximum a 16-bit linear value can represent. Wether the lowest f-stops really make sense or correspond to noise, will depend on the width of the DR, the shooting conditions and sensor quality.

This image (12MB) is the result of blending 3 shots at different exposures. Noise is practically negligible at all f-stops, and I achieved to record the almost 13 f-stops of the scene's dynamic range:




It also plots the EV distributions of any scene. For instance I used this image:



in an article about dynamic range, showing that noise is the real limitation of the DR in the shadows:


(nearly 8 f-stops of usable dynamic range in my 350D)

[Panopeeper]

showing that noise is the real limitation of the DR in the shadows

If one ignores noise, one can find details several stops under the "declared" bottom. One has to define, what is acceptable. The definition is easy, but it should be measurable, and that's not so easy.

Regarding the dynamic range, the logarithmic mode allows to calculate the dynamic range of a scene visually by using a log scale in the X axis

Calculating the range is one thing, but putting it on a correct scale is more difficult.

If you have a certain camera and you see a scaled histogram, you would expect that reflecting the camera. As the low end is not firm, one has to start at the clipping point, let's declare that 0EV or +10EV or whatever, and from there one can scale backwards. One has to "know the camera" in order to establish the 0EV point; the actual image tells nothing.

For example a fellow poster on DPReview shot a white board with the 40D *totally* overexposed (I don't know, with how many EV bias, but it must have been at least +3EV), so that all pixels are really clipped. He did this with all ISOs, and from that I established the list of clipping points. In the meantime I too got a 40D, which confirmed, that

I. there are variations between camera copies,

II. there are within a few levels, i.e. negligable,

III. the variations between ISOs are huge.

From here, one can draw an EV scale over the histogram.

It always plots 16 f-stops histograms, which is the maximum a 16-bit linear value can represent

I am often the domestic pedant (due to my education and my occupation). So, I have to remark: there is no such relationship.

This image (12MB) is the result of blending 3 shots at different exposures

This is an impressive what? Like a sports place, but there is nothing for the spectators. The same with theater. Perhaps an exhibition hall? (Or the inside of a paper box?)

[Guillermo Luijk]

If one ignores noise, one can find details several stops under the "declared" bottom. One has to define, what is acceptable. The definition is easy, but it should be measurable, and that's not so easy.

I agree, this is the same as we were previously commenting, noise can give information even if just the presence of noise reveals the darkest shapes.


Well my point may be simplistic but I think it is right, tell me if you agree: sensor is a linear device (of course not ideal, I did some tests to plot the real curve response and as one could expect in the ends: deep shadows and blown highlights, it starts to show non linear behaviour). So I set the 0EV reference in the 65534 level of the developed RAW, and start to count backwards allocating all levels in the image. E.g. a level 32768, in the middle of the range, would correspond to the border between f-stop 0EV and -1EV.
Levels with 65535 value are considered blown, i.e. higher than 0EV, so are ignored for the log histogram. In the same way I discard 0 values since they would belong to -infinite EV.

The non-linear behaviour in the ends could lead to some errors in the precise allocations of nearly blown highlights and deep shadows, but the middle part of the range (99% of the total information) should represent a real EV distribution of the scene under observation, don't you agree?

hehe that place is a memorial for the Alqaeda victims of the 11M in Madrid, so there is nothing to do there or to sit in but observe. Strangely shooting on a tripod is allowed in that place, while it is not in the rest of the train station. I waited till nobody was there since blending images with moving parts produces undesired ghosting (BTW I am improving my blending routine with an anti-ghosting algorithm that I think will work fine for small moving parts like tree leaves or water waves).
The room must be a square of about 20x20m with a strong natural light source in the middle of the ceiling which provoques a high dynamic range scene.

[John Sheehy]

I agree, this is the same as we were previously commenting, noise can give information even if just the presence of noise reveals the darkest shapes.
Well my point may be simplistic but I think it is right, tell me if you agree: sensor is a linear device (of course not ideal, I did some tests to plot the real curve response and as one could expect in the ends: deep shadows and blown highlights, it starts to show non linear behaviour).
[a href=\"index.php?act=findpost&pid=154902\"][{POST_SNAPBACK}][/a]

The shadows are totally linear in Canon RAW data, but they become non-linear when they are clipped at the blackpoint, snce all you have near black is a lot of zeros and some positive numbers due to noise, but all the negative numbers due to noise are clipped away.  The mean of black, after clipping, is always positive with noise.

This is why I think that blackpoint clipping should be saved until just before gamma is applied to the image.  Up to that point, offsets should be subtracted so that black is zero and there are negative numbers due to noise.

White balance, demosaicing, colorimetric adjustments, resampling, etc, can all be done with negative values present, and done better.

Think about what happens to the color of noise when the WB is done for incandescent - if you clip at black first, then zero is anchored, and blue noise gets exaggerated only in the positive direction.  If the blackpoint is unclipped, the WB pulls the negative blues lower, and the positive blues higher, and therefore the net, low-frequency effect is no exaggeration of the blue at all, and this is especially true if you downsample your image before blackpoint clipping.

[John Sheehy]

If one ignores noise, one can find details several stops under the "declared" bottom. One has to define, what is acceptable. The definition is easy, but it should be measurable, and that's not so easy.
Calculating the range is one thing, but putting it on a correct scale is more difficult.
[a href=\"index.php?act=findpost&pid=154844\"][{POST_SNAPBACK}][/a]

One of the issues that determines practical DR is desired resolution.  Many of the scientific approaches tend to overlook this aspect, and concentrate on the precision of individual pixels as measuring devices.  With a D3 or a 1Dmk3, You can probably read big, bold white fonts on a black background where white is 14 stops or more below RAW saturation, at ISO 1600.  I've done it 12 stops below with my 20D (white = 1 ADU, or about 0.8 average photons).  If the image were text with thin fonts, close to pixel width, then of course you have to take pixels and their noise more literally (no averaging possible), and would have relatively little usable DR.

[eronald]

Hello folks -
 Please allow me to hijack your interesting discussion a bit.
 I have a P45+ back. And am getting stripes in images from ISO 400.
 The back could give me good images up to ISO 3200 (I like grain), were it not for the stripes.
 Strange thing is, every Raw converter I've tried is writing *different* stripes !!!!
 I wonder whether this has something to do with the dark pixels.
 I can send people a link to file downloads of Raw images if interested.
 A glance here might be worthwhile if this topic interests you.
http://luminous-landscape.com/forum/index....showtopic=20901
 As I said, I can send people Raw images or make some; the backs are well linearised, I believe and have good A/D converters.

Edmund

[Panopeeper]

I have a P45+ back. And am getting stripes in images from ISO 400

Well, Edmund, you are one of the few lucky ones, who have not been confronted with this phenomenon before (those, who say otherwise, have not taken a really close look and they don't know, that this is omnipresent).

Anyway, instead of speculation, what about posting the raw images, so that others can take a really close look?

[Guillermo Luijk]

yes, I also would like to take a look at your RAW files, even at ISO400 you could have underexposed the escene and there is no sensor today that won't display some noise in an underexposed ISO400 image.
Also there seems to be some trouble with the blue channel: I think the black point of the blue channel is wrong and that gave your pic a blue appearance and a lot of blue noise in the shadows.

[Panopeeper]

tell me if you agree: sensor is a linear device (of course not ideal, I did some tests to plot the real curve response and as one could expect in the ends: deep shadows and blown highlights, it starts to show non linear behaviour)

Sensors are very linear devices, perhaps even more, than you think.

Pls elaborate the circumstances you think, that they are not linear in the highlights (the shadow is a different issue, for it is difficult to draw the line).

I have seen demonstrations of non-linearity with Imatest, which turned out to be useless, at a close look.

So I set the 0EV reference in the 65534 level of the developed RAW, and start to count backwards allocating all levels in the image

The numerical range of the raw values may be {black level - 4095} or {black level - 16383} or {black level - 65534} etc. or anything in between.

What is your basis to put the 0EV at any particular point, if you don't know that particular camera?

For example the Nikon D2X tops at 3880 with greens, while the reds go to 4095I listed already the pecularities of the Canon 40D.

This depends on ISOs and on the "kinfd" of the p[ixel in the CFA.

Even worse: the two greens of the CFA (remember: red-green, green-blue, this is the most popular arrangement) can have different responses.

I reiterate it: in order to establish an EV scale, one has to be able to define a known, proven point. In other words: if you don't know the camera's characteristics (at all ISOs), then you can not scale the histogram. If you do, that is good only for the "unwashed".