[lurchlarson]

I have shot film and I agree that my K200d is way cleaner at all comparable ISOs than film was. One thing that people often forget is that sometimes noise actually ADDS to the artistic quality of the image.

But all in all. I think the biggest deal is that everyone wants to ride the "gripe-train" to feel cool and to feel like they have something important to say about technical stuff. I've been guilty of that myself and I think most have at some point. That's not to say that I wouldn't like having noise-free images at 6400. I know that that is still a few years away though.

I've found that my photographic and artistic skill has increased by figuring out how to make a scene work when I am in high-iso situtaions.

[simico]

Imo those people writing the "high ISO suxx" posts

• expect too much from high ISO and have no clue how better it is than X years ago
• can't even make a high ISO shot with proper exposure and settings (at least most of the posted "supporting pictures" are (badly) underexposed and/or not even properly focused, etc)

Properly made shots and realistic expectations help a lot.

[Lowell Goudge]

maybe it's time for a poll. well defined (of course) to check out the expectations of people. the poll needs to ask about acceptance of todays cameras, but should also ask questions about when you started shooting.

It seems that my initial idea about 2 groups seems to be playing out.

[newarts]

I think high iso performance is running into a few fundamental barriers, the most important of which is physics.

The Signal/Noise ratio of an exposed pixel is proportional to the square root of the number of photons collected during the exposure. This number depends *only* on the intensity of the incoming light and the area of the collecting pixel (neglecting so-called dark noise.)

Say you had a sensor pixel with a really high quantum efficiency - three electrons were collected for each incoming photon - while this sounds great it would not change the inherent S/N ratio which is fixed by the incoming light and is not influenced by any subsequent constant multipliers.

Given a projected 2D image with particular light intensity distribution and exposure time, the *only* way to increase the S/N ratio per pixel is to use bigger pixels. Period.

But as pixel size grows, geometric resolution decreases; there is an inherent balance between the two and there is nothing one can do about it even with perfect optics.

S/N = PixelSize*Sqrt(Intensity*time)
Resolution = Constant/PixelSize

Therefore, for a particular scene and exposure time:

S/N = AnotherConstant/Resolution - as resolution gets better, noise gets worse.

Mathematical noise reduction can help a lot; say you are taking a photo of something like poster that consists of areas of uniform color; in principle all noise can be eliminated from such a photo since you know a-priori that the result must be noiseless areas with perfectly sharp boundaries.

Unfortunately, natural scenes contain natural noise, so there cannot be a perfect way to get rid of all the noise that's "not supposed to be there" if you don't know beforehand what's not supposed to be there; I'm thinking of a photo of a pile of dirt for example. (nb; the light's noise varies with time, while the dirt-pile's noise varies with position - therefore they are separable to some degree.)

I think at this point in the development of digital photography desire for noiseless photos is mostly in the capable hands of people writing noise reduction software that takes typical scene properties and human visual characteristics into account.

The light from the sky on a bright day is inherently noisy if observed for a short enough time. A human observer's mind however "knows" the sky isn't noisy - this type problem can be mitigated in software.

Dave

PS This is only one important aspect of the situation, the other being noise that doesn't depend on the incoming light (also getting lower as the technology gets better.)

I hope I'm wrong about these limits but don't think so.

---

Last edited by newarts; 07-13-2009 at 07:01 AM.

[jstevewhite]

I think high iso performance is running into a few fundamental barriers, the most important of which is physics.

<snip>

I hope I'm wrong about these limits but don't think so.

I think you're right as far as your post goes; bayer sensors and current technologies must depend on NR software to reduce noise. However, you're assuming that technologies don't change. Temporal noise is susceptible to frame averaging, for instance. If we push the sensors enough that we can use the actual sensors as shutters (at lower speeds we could already), then a camera could make, say, four 1/4000 second exposures for a 1/1000 final, average them, and virtually eliminate temporal noise. Might be a few years off, but I think it will come; we'll probably see it as an option first, where you can turn on "MEA" (multiple exposure averaging) on high-end DSLRs, and it will work at a limited exposure span to start with, like "below 1/250 sec". You can approach something similar with multiple exposures on the current K20D, for tripod shots.

That's just one way I've read that researchers are pursuing noise reduction in smaller (sub-5-micron) sensels. There are others that I've read about, but don't immediately spring to mind.

[2Thumbs]

In certain siuations I find noise can add to a photograph. Such as in band photography at some seedy back alley club (or rather, if that's the portrayal you want to give accross to the viewer). It also makes for an excellent 'film grain' effect when used in b&w.

Fortunatly the 20D has excellent noise handling so 1600 is very usable, 3200, on the over hand, isn't though in my experience.

[jstevewhite]

Imo those people writing the "high ISO suxx" posts

• expect too much from high ISO and have no clue how better it is than X years ago
• can't even make a high ISO shot with proper exposure and settings (at least most of the posted "supporting pictures" are (badly) underexposed and/or not even properly focused, etc)

Properly made shots and realistic expectations help a lot.

Hear, hear. I've been absolutely amazed at the high-iso performance of these cameras with "poor high-iso performence".

Hey, I get it, though. Just because it looks better than old stuff doesn't mean it couldn't look any better. It's like hot days in the midwest. On a 90 degree July day with 70% relative humidity, if you complain about the heat, someone will point out that it's gonna get hotter before it gets cooler: "This isn't that hot - just wait 'til August!" Just 'cuz it's gonna get (or has been) worse, doesn't mean this it's *good* right now.

In the interim, though, I think it behooves us to note that we've all see beautiful images shot on 35mm in Tri-X pushed to 800, 1600, and occasionally higher. We talk about the technical bits of photography, of course; most of us are gear-heads. But we should remember that technique and equipment != beauty. A picture of a brick wall shot with the best system in the world is probably going to be less interesting than a light-fogged, unsharp personal portrait done with a plastic Holga. What you point your camera at is probably at least as important as what kind of camera it is.

[Lowell Goudge]

I am going to disagree with some of what you have below, but I am not picking on you just making sure people are all clear on a few things.

I think high iso performance is running into a few fundamental barriers, the most important of which is physics.

I agree, but not everyone understands that by wishing for something you can't break the laws of physics

The Signal/Noise ratio of an exposed pixel is proportional to the square root of the number of photons collected during the exposure. This number depends *only* on the intensity of the incoming light and the area of the collecting pixel (neglecting so-called dark noise.)

not really, there are multiple aspects to noise, what you have defined here is resolution. i.e. what is the change impact of an additional electron. Dark noise cannot be ignored, and this is exactly why bigger sensors have less noise. Dark noise is always present, but the impact of a single electron (as you put it) in a bigger sensor is lower, this is part of why the P&S cameras have such bad noise, because they have such small sensors dark noise becomes a bigger and bigger part of it. You also have the amplifier chain with it's own induced errors that may change from sensor to sensor as data is collected.

Say you had a sensor pixel with a really high quantum efficiency - three electrons were collected for each incoming photon - while this sounds great it would not change the inherent S/N ratio which is fixed by the incoming light and is not influenced by any subsequent constant multipliers.

again you are ignoring the subsequent multipliers, they are not perfect.

Given a projected 2D image with particular light intensity distribution and exposure time, the *only* way to increase the S/N ratio per pixel is to use bigger pixels. Period.

again you are mixing noise with resolution.

But as pixel size grows, geometric resolution decreases; there is an inherent balance between the two and there is nothing one can do about it even with perfect optics.

S/N = PixelSize*Sqrt(Intensity*time)
Resolution = Constant/PixelSize

Therefore, for a particular scene and exposure time:

S/N = AnotherConstant/Resolution - as resolution gets better, noise gets worse.

I agree, but you need to clarify optical resolution

Mathematical noise reduction can help a lot; say you are taking a photo of something like poster that consists of areas of uniform color; in principle all noise can be eliminated from such a photo since you know a-priori that the result must be noiseless areas with perfectly sharp boundaries.

this is called salt and pepper noise in PSPX2 which I use

Unfortunately, natural scenes contain natural noise, so there cannot be a perfect way to get rid of all the noise that's "not supposed to be there" if you don't know beforehand what's not supposed to be there; I'm thinking of a photo of a pile of dirt for example. (nb; the light's noise varies with time, while the dirt-pile's noise varies with position - therefore they are separable to some degree.)

I think at this point in the development of digital photography desire for noiseless photos is mostly in the capable hands of people writing noise reduction software that takes typical scene properties and human visual characteristics into account.

The light from the sky on a bright day is inherently noisy if observed for a short enough time. A human observer's mind however "knows" the sky isn't noisy - this type problem can be mitigated in software.

Dave

PS This is only one important aspect of the situation, the other being noise that doesn't depend on the incoming light (also getting lower as the technology gets better.)

I hope I'm wrong about these limits but don't think so.

[slrl0ver]

newarts,

I think you're talking about SNR but not taking into consideration the complexities of the semiconductors we use as photon gathering devices and the associated "signal chain" to convert from the analog domain into the digital.

I suspect the CMOS/CCD sensor design, A/D's, and power supply conditioning are are more of a limiting factor for noise performance than the available light for sensors < 20 Mpixel in sub-35mm sensors. It's why other manufacturer's have continued to improve noise and low-light performance WITHOUT sacrificing details (i.e. filtering). But, some of those companies are reaching limits which I why I speculate the floor is around 20 Mp.

I don't think that's a startling observation given how long the respective companies have been involved in the design of their own semiconductor sensors.

-slrl0ver

I think high iso performance is running into a few fundamental barriers, the most important of which is physics.

The Signal/Noise ratio of an exposed pixel is proportional to the square root of the number of photons collected during the exposure. This number depends *only* on the intensity of the incoming light and the area of the collecting pixel (neglecting so-called dark noise.)

Say you had a sensor pixel with a really high quantum efficiency - three electrons were collected for each incoming photon - while this sounds great it would not change the inherent S/N ratio which is fixed by the incoming light and is not influenced by any subsequent constant multipliers.

Given a projected 2D image with particular light intensity distribution and exposure time, the *only* way to increase the S/N ratio per pixel is to use bigger pixels. Period.

But as pixel size grows, geometric resolution decreases; there is an inherent balance between the two and there is nothing one can do about it even with perfect optics.

S/N = PixelSize*Sqrt(Intensity*time)
Resolution = Constant/PixelSize

Therefore, for a particular scene and exposure time:

S/N = AnotherConstant/Resolution - as resolution gets better, noise gets worse.

Mathematical noise reduction can help a lot; say you are taking a photo of something like poster that consists of areas of uniform color; in principle all noise can be eliminated from such a photo since you know a-priori that the result must be noiseless areas with perfectly sharp boundaries.

Unfortunately, natural scenes contain natural noise, so there cannot be a perfect way to get rid of all the noise that's "not supposed to be there" if you don't know beforehand what's not supposed to be there; I'm thinking of a photo of a pile of dirt for example. (nb; the light's noise varies with time, while the dirt-pile's noise varies with position - therefore they are separable to some degree.)

I think at this point in the development of digital photography desire for noiseless photos is mostly in the capable hands of people writing noise reduction software that takes typical scene properties and human visual characteristics into account.

The light from the sky on a bright day is inherently noisy if observed for a short enough time. A human observer's mind however "knows" the sky isn't noisy - this type problem can be mitigated in software.

Dave

PS This is only one important aspect of the situation, the other being noise that doesn't depend on the incoming light (also getting lower as the technology gets better.)

I hope I'm wrong about these limits but don't think so.

[spartan]

High ISO Shots - PENTAX K-7 VS NIKON D300 [Page 2]: Pentax SLR Talk Forum: Digital Photography Review

True. However, the fact remains that based upon teh same measurement system the pentax sensor is rated at a 15%-20% higher sensitivity than the Nikon. At this point in the ISO curve, a small change in ISO can bring about a large noise level.

[Rondec]

I just don't think people understand the relationship between exposure and high iso noise. Digital just cannot be pushed in the same way that film can. If you push digital, even at lower iso settings, you tend to introduce noise. This was true in the 6 MP cameras like the K100 and it is even more true with the K20. The camera cannot save bad photographers from themselves.

The other thing that is often over looked is that the size of printing effects how noticeable noise is. I routinely shoot at iso 1600 on the K20. There may be some noise, but at the size I'm printing (8x10 or less), it is not noticeable, whereas at 100 percent on a monitor, it is.

[Lowell Goudge]

The other thing that is often over looked is that the size of printing effects how noticeable noise is. I routinely shoot at iso 1600 on the K20. There may be some noise, but at the size I'm printing (8x10 or less), it is not noticeable, whereas at 100 percent on a monitor, it is.

I was completely happy with my K10D and noise until I purchased a 22 inch wide format monitor. The viewable section is 13 x 19 inch. quite a step up from my 4/3 17 inch monitor.

we do tend to pixel peep to the nth degree today, and I have printed crops out of a K10 onto 11 x 17. The monitors are so far advanced compared to printing, that noise is not a big issue.

[jstevewhite]

I was completely happy with my K10D and noise until I purchased a 22 inch wide format monitor. The viewable section is 13 x 19 inch. quite a step up from my 4/3 17 inch monitor.

we do tend to pixel peep to the nth degree today, and I have printed crops out of a K10 onto 11 x 17. The monitors are so far advanced compared to printing, that noise is not a big issue.

Well, the dot pitch of monitors is so low that the noise becomes obvious at 100%; Screens run 72-100 DPI (usually), while printers are generally at least 300dpi for photo use. But you're right - I've printed ISO1600 shots at 8x10 and they're gorgeous.

[Haakan]

While it seems appealing to be able to have the image quality of a camera represented by a simple number such as the DxO Mark, I am personally not convinced that we are there yet, especially looking at their low noise measure.

First, even while DxO produces a lot of measured values, it seems as there is quite a lot of subjectivity introduced when translating these values into composite numbers by the introduction of thresholds that has been set subjectively.

As one of the components in their low noise measures they use a threshold for the dynamic range, but IMHO their measure does not seem capture what is perceived as dynamic range when looking at a picture. While it would be nice, I do not think that an APS sensor camera (Nikon D90), at ISO 200, produces images on par with a 39 MPix Hasselblad MF format camera at ISO 50 in terms of how one perceive the dynamic ranges of an image. I think that a part of the issue is that the DxO score for dynamic range only look at the noise in the darkest area, and not if the darker area in itself is distinguishable to the next lesser darkest area. I.e. is it relevant to talk about an increase in dynamic range, if the extension of the range does not produce a distinguishable tonality? If taken to the extreme, a sensor which maps all signals below a certain threshold to pure black will have a very good SNR at the black areas, but can that be seen as a good measure to determine dynamic range?

For me, when looking at low ISO performance the image quality is much more related to the overall noise in the whole image, across all tonal ranges (this is the SNR score in DxO Mark), and to a lesser extent to the noise in the extreme darkest part of the image. Especially if the lower noise in the darkest area comes at the expense of not being able to separate out different shades in the deep darkness. I.e. can you claim a larger dynamic range if two nearby shades in the darkest areas are not possible to separate by tonal difference?

Since the DxO Marks for low noise performance of D90 compared to K20D and EOS50D has triggered quite a lot of discussions in different fora, I thought I would try to make my own comparison of the high ISO image quality using RAW files from Imaging Resource.

So I downloaded RAW files from Imaging Resource for K20D (K7 not yet available), D90 and EOS50D. I then converted them all in Raw Therapee, using "zero" settings for all values, except for exposure which was adjusted to give all images a similar white level. I used was the "multi targets" from Imaging Resource from which I cropped out the "Kodak Target" as a good representative image.

I have attached JPEGs of the resulting images below, scaled to the same sensor resolution (12 Mpix), 100% crops. When looking over the images, there seems to be some inconsistencies in the focusing, but since I only look at the noise structure, that should not be an issue.

When I looked at the resulting images, this is my take on them (note that this is subjective and others might come to other conclusions).

1. All 3 cameras are extremely capable and as indicated by the DxO SNR curves, they are all very close in overall "image noisiness", with a slight advantage to K20D at lowest ISO, in having 1 dB lower noise at its lowest ISO vs. the D90. I.e. at its best setting the K20D should produce a slightly cleaner overall images than the D90. But at higher ISO, there is a slight advantage to the D90.

2. Looking at the higher ISO images (ISO 1600 and 6400) for all tonal ranges, except for the darkest tones, there seems to be no big difference in the noise levels. To me they all give a similar "noise impression". The D90 looks cleaner in the darkest tones whereas the K20D looks cleaner in the lighter tones. Overall the EOS50D looks slightly noisier but the difference is quite small.

3. Looking at the character of the noise, I personally feel that the noise for K20D is somewhat smoother looking than for the D90. The D90 seems to have more "posterized" noise. When I cleaned up the images in "Neat Image" I felt I liked the result from K20D better than for D90. But the differences are really small and only found when pixel peeping. And I realize that since I have a Pentax, I might be somewhat biased in my finding. With the differences being quite subtle, I think that Nikon and Canon owners would be equally right in claiming a preference for their respective brands.

So, overall I feel that the DxO Mark, while a good attempt to try to help out in benchmarking performance, has not yet been able to fully correlate their numbers to how I perceive image quality, especially true for the composite numbers. And for some people it might be confusing when the presented numbers are a mix of logarithmic and linear numbers. E.g. color depths are given in bits (or EV), whereas the low light performance is given in ISO which is a linear scale. If converted to e.g. EV (and use a base ISO to put the number in a similar range as the other numbers), the low light ISO values for K20D, D90 and EOS50D could equally well be given as 19.3, 19.9 and 19.4 respectively (which might not be perceived as so dramatic differences as their linear representations). And these numbers are still very heavily influenced by the way DxO measures dynamic range, and not so much related to the overall all noise in the image.

So IMHO, despite the sometimes raging online wars on which camera of these three is the better, to me the differences looks quite minor at least in the above comparison, despite their difference in DxO scores

Now I have spent far too much time on this project, and will try to shoot some nice photos instead.

Best regards,
Haakan

K20D ISO 100


D90 ISO 200


K20D ISO 1600


D90 ISO 1600


EOS50D ISO 1600


K20D ISO 6400


D90 ISO 6400


EOS50D ISO 6400

[Ash]

Great contribution to the discussion Haakan,
*Real life* results are more important than what lab testers report.
We can pixel peep all we like, and some of us won't be satisfied until we have noiseless ISO 40,000 images, but photography is a skill in technique and understanding how light works, which we want to replace with gear that would take care of all that for us.

How far do we want to push AF and high ISO noise barriers? Until we can take a sharp handheld photo in pitch blackness with no flash?