[richardrosenman]

Hi gang;

I am aware of the different noise models usually found in digital photography such as Uniform, Gaussian, Laplace, Lorentz, Poisson, etc.

But do photographs usually contain various types of noises combined or just one?

I've read noise removal software depends on reading the image's histogram to find out the noise distribution and thereby identify which type of noise is in an image but this suggests there is always only one type of noise inherent?

I can't seem to find any information on this. Any info is appreciated.

-Rich

[clackers]

Two part article here:

Digital Camera Image Noise: Concept and Types

[richardrosenman]

Two part article here:

Digital Camera Image Noise: Concept and Types

Thanks for your reply.

Yes, I have seen this article and gone through it. The article explains three different types of camera noise but it doesn't discuss if these three can be found together in one image. Rather, it seems to suggest that an image will consist of one of these types of noises.

I am trying to ascertain if digital photographic images contain several types of noises together in one, or only ever one type.

-Rich

[sjwoodworth]

The different types of noise are covered in quite some depth in the April 20th lecture of Marc Levoy's online course:
Course schedule - Digital Photography

(I don't know the material well enough to attempt providing the informed answer myself, though yes, there are different sources and types of noise!)

---

Last edited by sjwoodworth; 02-15-2017 at 08:16 PM.

[photoptimist]

For the most part, the noise in the RAW image pixels is additive so that the measured pixel value is the true value plus a bunch of different types of noise. There's a type of noise that is proportional to the square root of the linear pixel value which can add or subtract from the "true" value, but subtraction is a type of addition.

The histogram can help but only if you know what the histogram is supposed to look like if the sensor had no noise (e.g., image with the lens cap on should have a histogram with a perfect spike at "zero").

[stevebrot]

The article explains three different types of camera noise but it doesn't discuss if these three can be found together in one image. Rather, it seems to suggest that an image will consist of one of these types of noises.

After having researched how each type of noise is generated and their characteristics, ask yourself this question,

"Why would they not occur together?"

Can you possibly think of any reason why one type would exclude the other? Alternatively, you could simply look at a sampling of your own images. The Cambridge in Color article linked by @clackers lays it out in fairly simple terms that are directly observable in one's own images. Now for the inevitable set of bullet points:

• There are only two types of noise: Luminance and Color (Chroma)
• The other classifications (Gaussian, Poisson, Uniform, etc.) describe the distribution and/or the pattern and nothing more
• Every image on my hard drive has both luminance and color noise. Their presence and type can be demonstrated by using the noise reduction sliders in Lightroom.
• All noise is a replacement of data, on a pixel-by-pixel basis, with non-data and the distribution is not truly random. That simple fact has huge ramifications for noise reduction actions. When one removes the noise, what is it replaced with (ponder that one) and what does that mean in regards to detail capture? Detail is organized data and when the data (detail) was never captured in the first place, how does one make everything look OK without both the data and the distraction of that pesky noise.

P.S. In case it was missed the first time around: Cambridge In Color: Image Noise Tutorial, part 1 (The second part is the better of the two.)


Steve

[lithedreamer]

I think about noise the way Lightroom thinks about it, since that's where I'm usually working. In that model, there's luminance noise and colour noise. The latter is definitely more distracting, as it usually follows luminance noise and produces splotches of colour all over an image.

Here's a random example I found:



Some moderate colour noise can be fairly easily removed in post if you're shooting raw, as can quite a bit of luminance noise. On the other hand, luminance noise in reasonable amounts feels to me like 'digital grain', and can add to an image sometimes.

I'm sure others can answer this better, but I just wanted to provide an in-thread answer.

[Sagala]

Hi gang;

I am aware of the different noise models usually found in digital photography such as Uniform, Gaussian, Laplace, Lorentz, Poisson, etc.

But do photographs usually contain various types of noises combined or just one?

I've read noise removal software depends on reading the image's histogram to find out the noise distribution and thereby identify which type of noise is in an image but this suggests there is always only one type of noise inherent?

I can't seem to find any information on this. Any info is appreciated.

-Rich

Where do you read this: ''noise removal software depends on reading the image's histogram''?

[richardrosenman]

Where do you read this: ''noise removal software depends on reading the image's histogram''?

It was an interesting find actually. I will have to try and it find it but this seems to be a common concept for noise removal software as I saw it mentioned in more than one reference.

They say that because these noise distributions have specific function curves, to identify what type of noise distribution your image has, you simply look at the histrogram. The histrogram of the noise should look like one of these distribution curves:



Source: http://www.slideshare.net/kalyanacharjya/image-restoration-40589017

Once you can identify the distribution type, you are then one step closer to removing noise successfully by knowing what algorithm to work with.

I, on the other hand, am actually interested in noise generation for images - thus my original question. It would make sense that several types of noises are present in any photographic image. I suppose I was asking more about the noise type (distribution) to create real-world noise. For instance, I have also read that uniform noise rarely occurs in nature and is highly unreaslistic whereas gaussian noise is a much more natural type of noise.

@lithedreamer; Are you sure that both color noise AND luma noise are found together? In the example you posted, it seems like luma noise is created (and left over) after your software attempted to reduce the color noise. It would seem it tried to normalize the colors thereby leaving their luminance intensities which show up as luma noise... no? You brought up a very interesting point - I never considered that you can have color noise AND luma noise together so I would like to know the answer to this as well.

Thanks for your replies guys - please keep them coming. Any insight into other characteristics of photographic image noise is worth mentioning here.

EDIT: Here is where I found that discussion regarding identifying noise type by looking at the histogram: Link

-Rich

[Sagala]

It was an interesting find actually. I will have to try and it find it but this seems to be a common concept for noise removal software as I saw it mentioned in more than one reference.

They say that because these noise distributions have specific function curves, to identify what type of noise distribution your image has, you simply look at the histrogram. The histrogram of the noise should look like one of these distribution curves:



Source: Image restoration (Digital Image Processing)

Once you can identify the distribution type, you are then one step closer to removing noise successfully by knowing what algorithm to work with.

I, on the other hand, am actually interested in noise generation for images - thus my original question. It would make sense that several types of noises are present in any photographic image. I suppose I was asking more about the noise type (distribution) to create real-world noise. For instance, I have also read that uniform noise rarely occurs in nature and is highly unreaslistic whereas gaussian noise is a much more natural type of noise.

@lithedreamer; Are you sure that both color noise AND luma noise are found together? In the example you posted, it seems like luma noise is created (and left over) after your software attempted to reduce the color noise. It would seem it tried to normalize the colors thereby leaving their luminance intensities which show up as luma noise... no? You brought up a very interesting point - I never considered that you can have color noise AND luma noise together so I would like to know the answer to this as well.

Thanks for your replies guys - please keep them coming. Any insight into other characteristics of photographic image noise is worth mentioning here.

EDIT: Here is where I found that discussion regarding identifying noise type by looking at the histogram: Link

-Rich

I was confused because image histogram and noise histogram is not the same thing.
Photographic noises generators sofwares exist (DxO FilmPack ,etc.).You can also samples a noise patch in a photo and use it with a seamless textures software like Hypertyle to apply noises as you wish. Sorry for my english.Sagala.

[photoptimist]

The issue of chrominance noise is actually more an artifact of the RAW development software than it is true noise in the data.

If one looks at just the red channel of a Bayer filter sensor output, only a 1/4 of the pixels detect red light. By the same token, the red pixels have no data on the amount of green and blue light on those pixels. It's the demosaicing software (either the JPEG engine in the camera or the RAW development software in the computer) that attempts to fill-in the remaining 75% of missing red pixels (and 75% of missing blue pixels, and 50% of missing green pixel). Compounding this issue is the fact that digital sensors tend to be somewhat less sensitive to red and blue light relative to green light which forces the software to amplify red and blue.

The demosaicing software is forced to make a lot of assumptions about what's happening in all those missing pixels. Is the picture an aquarium full of colored beads, a garden full of flowers, a page of newspaper text, a smooth sunset sky? Depending on the scene, the "correct" values for all the missing pixels will be different and the discrepancy between what the software filled in and what was truly there (aka the "noise") will be different.

In short, chrominance noise and luminance noise do become coupled because the software must make a choice: are individual variations in red, green, and blue pixels a sign that the color is varying a lot or are they a sign that luminance is varying a lot? Depending on the subject matter, either assumption might be true. In fact, different parts of the image might be different.

[richardrosenman]

The issue of chrominance noise is actually more an artifact of the RAW development software than it is true noise in the data.

If one looks at just the red channel of a Bayer filter sensor output, only a 1/4 of the pixels detect red light. By the same token, the red pixels have no data on the amount of green and blue light on those pixels. It's the demosaicing software (either the JPEG engine in the camera or the RAW development software in the computer) that attempts to fill-in the remaining 75% of missing red pixels (and 75% of missing blue pixels, and 50% of missing green pixel). Compounding this issue is the fact that digital sensors tend to be somewhat less sensitive to red and blue light relative to green light which forces the software to amplify red and blue.

The demosaicing software is forced to make a lot of assumptions about what's happening in all those missing pixels. Is the picture an aquarium full of colored beads, a garden full of flowers, a page of newspaper text, a smooth sunset sky? Depending on the scene, the "correct" values for all the missing pixels will be different and the discrepancy between what the software filled in and what was truly there (aka the "noise") will be different.

In short, chrominance noise and luminance noise do become coupled because the software must make a choice: are individual variations in red, green, and blue pixels a sign that the color is varying a lot or are they a sign that luminance is varying a lot? Depending on the subject matter, either assumption might be true. In fact, different parts of the image might be different.

Ok - but it is impossible to visually see both chroma and luma noise in an image right? Because the chroma noise would mask the luma noise. So we are only assuming that both can co-exist? Is there any way to know this for sure.

Regarding the noise distribution - do we know which distribution types exist in photographic noise? Everyone points to gaussian but they say Lorentz (salt and pepper noise) also exists. Do we have a definitive list of what those would be?

Thanks

[rechmbrs]

Ok - but it is impossible to visually see both chroma and luma noise in an image right? Because the chroma noise would mask the luma noise. So we are only assuming that both can co-exist? Is there any way to know this for sure.

Regarding the noise distribution - do we know which distribution types exist in photographic noise? Everyone points to gaussian but they say Lorentz (salt and pepper noise) also exists. Do we have a definitive list of what those would be?

Thanks

What's been mentioned so far are noise types related to the sensor. There is a great amount of what might be called light travel path related noises. Reflections, refractions, diffractions, other light sources and on and on. They most often are hard to see and near impossible to suppress and obscure the desired signal. A bright street light either is signal or noise depending on the scene. These noises are significantly stronger than the sensor noise but because of the difference in character the sensor noise is suppressive. Software is often our worst enemy as far as accentuating some noises. Sharpening and the bright street (spike like features of any sort) can destroy your masterpiece.

Enough from me.
RONC

[richardrosenman]

Thanks Ron, for that. It's helpful.

Ok, so let me summarize what we've gone through so far.

A capable noise generation software should:

- be able to produce chromatic noise AND luminance noise TOGETHER so for instance chromatic gaussian noise combined with Lorentz (Salt and Pepper) luminance noise (I am still unconvinced about this but...)
- be able to apply noise using the Uniform, Gaussian, Laplace, Lorentz, and Poisson distributions which seem to be the main photographic types of noise distribution (any others? Gamma? Exponential? Rayleigh?)
- be able to apply these different distributions TOGETHER so a combination of standard gaussian noise with Lorentz (Salt and Pepper) noise
- be able to control each type of noise distribution intensity separately (so that the level of gaussian noise may be different than the level of Lorentz (Salt and Pepper) noise
- be able to produce photometric noise? (we didn't discuss this - the application of noise more on the darker parts of an image that receive less light)

Does everyone agree?

What else would you guys add?

What other features might be handy?

-Rich

[clackers]

Thanks Ron, for that. It's helpful.

Ok, so let me summarize what we've gone through so far.

A capable noise generation software should:

- be able to produce chromatic noise AND luminance noise TOGETHER so for instance chromatic gaussian noise combined with Lorentz (Salt and Pepper) luminance noise (I am still unconvinced about this but...)
- be able to apply noise using the Uniform, Gaussian, Laplace, Lorentz, and Poisson distributions which seem to be the main photographic types of noise distribution (any others? Gamma? Exponential? Rayleigh?)
- be able to apply these different distributions TOGETHER so a combination of standard gaussian noise with Lorentz (Salt and Pepper) noise
- be able to control each type of noise distribution intensity separately (so that the level of gaussian noise may be different than the level of Lorentz (Salt and Pepper) noise
- be able to produce photometric noise? (we didn't discuss this - the application of noise more on the darker parts of an image that receive less light)

Does everyone agree?

What else would you guys add?

What other features might be handy?

-Rich

I'm not sure you've thoroughly read the Cambridge articles, if you say chroma and luminance noise can't coexist.

The shot noise in the sensor's pixels has a Poisson distribution, for instance.