Noise patterns

[churly]

Does anyone know of any public domain studies that characterize noise patterns for digital sensors. 

I make a lot of handheld, rather high ISO, shots on my crop sensor D7200 and am thinking about having a go at writing some code, probably using MATLAB, to work on noise.

In general the various noise removal applications that I have tried work quite well when the shot is properly exposed to the right but in the case of less well exposed images, the noise removal filters still leave distinctive, low level, unpleasant noise patterns (unless you go for the plastic look).  I would like to look into the structure of the left-over noise and have a go at dealing with it.  Presumably the noise patterns are a function of the specific sensor and the algorithm used deal with the raw file so I'm thinking to customize something for my specific purposes.

Any thoughts would be welcome.

Chuck

[kirkt]

I do not know of such a repository, but if you are finding specific pattern noise (like all of my Canon 5D series cameras had when I shot with them) then you may want to try pre-filtering for the pattern with a FFT approach.  I was able to remove the pattern line noise from each grayscale channel and then recombine the channels so that only the "random" noise remained - this is what most NR applications deal with and I could use these applications (I used Neat Image) without worrying about the line pattern.  Topaz DeNoise (the old application, not the new AI-driven one) had a feature that would also attempt to remove this pattern (horizontal or vertical lines) but it was sort of hit-or-miss.  FFT filtering is good for removal of patterns that are regular and repetitive.  I have also used it for other applications as well, for example, removing the hex cell pattern that often is overlaid on old photo prints by the photo album sticky page covers.

You can also make constructing the FFT filter a little bit easier for certain patterns if you slightly rotate the image so that the pattern is not parallel to the vertical or horizontal edge of the image.  This will make the pattern easier to isolate from the axes of the FFT spectrum that contain the actual image data.

A good tool for prototyping the FFT filters is NIH Image/ImageJ/Fiji - all the same basic open, free image processing application.  It also has a macro programming language.

http://fiji.sc

Good luck!

kirk

EDIT - here is an image that has a pattern overlaid on top of grain.  It is analogous to patterned noise overlaid on random noise.  You can see how the FFT filtering has removed the pattern but left the random noise intact for separate treatment.  This way you can gently address the random noise to your liking without having to hammer it over the head to try to remove the pattern too.  BTW, Affinity Photo has live FFT filtering built into it as a regular filter.

[churly]

Kirk - Thanks much for the info.  I don't have issues with pattern noise it is more a question of whether the 'random noise' is truly random and how to deal with the residual after filtering.  I haven't done anything rigorous yet but my impression is that there is a small-scale structure in the 'random' noise.  I need to dig in deeper but I was just wondering if others have seen any literature on this.  Yes, I would expect to do the filtering in the frequency domain - quick computation and easy design.

Many thanks for your insight!

Chuck

[sbay]

I don't have any studies for you but have you looked at what is already done in astrophotography for removal of noise, including fixed pattern noise? There's a lot of information available on that.

[churly]

Thanks sbay.  Yes, I have just started dipping into the astrophotography literature and as you say there is a lot of info.  It has definitely helped me to better define the problem.

[Jim Kasson]

Kirk - Thanks much for the info.  I don't have issues with pattern noise it is more a question of whether the 'random noise' is truly random and how to deal with the residual after filtering.  I haven't done anything rigorous yet but my impression is that there is a small-scale structure in the 'random' noise.  I need to dig in deeper but I was just wondering if others have seen any literature on this.  Yes, I would expect to do the filtering in the frequency domain - quick computation and easy design.

Many thanks for your insight!

Chuck

You're using Matlab, right? Just do an FFT on a dark frame, and you'll see the periodic noise.

https://blog.kasson.com/?s=read+noise+spectra

Jim

[Guillermo Luijk]

I don't have issues with pattern noise it is more a question of whether the 'random noise' is truly random

Sure it is. What you call 'random noise' in a digital sensor is produced by thermal effects on the electronics. It is known as 'read noise' and follows a gaussian distribution. There is actually no difference between one sensor or another regarding read noise as to speak of differentiated noise signatures.

People usually confuse the differentiated appearence of noise grain after RAW development, which is caused basically by the software process applied, with particular kinds of noise among different sensors. There is no such difference, all of them are gaussian with zero mean on every single RAW channel and mostly incorrelated with neighbour photosites.

Regards

[Jim Kasson]

Sure it is. What you call 'random noise' in a digital sensor is produced by thermal effects on the electronics. It is known as 'read noise' and follows a gaussian distribution. There is actually no difference between one sensor or another regarding read noise as to speak of differentiated noise signatures.

People usually confuse the differentiated appearence of noise grain after RAW development, which is caused basically by the software process applied, with particular kinds of noise among different sensors. There is no such difference, all of them are gaussian with zero mean on every single RAW channel and mostly incorrelated with neighbour photosites.

I've seen huge differences among sensors wrt the spectral distribution of the 4 raw channel read noise. Local heating causes low frequency bumps. PDAF processing causes periodic ones (eg PDAF banding). You can sometimes see where the stepper exposed the chip nonuniformly. The G1 and G2 channels often have different spectra.

Jim

[Doug Gray]

I've seen huge differences among sensors wrt the spectral distribution of the 4 raw channel read noise. Local heating causes low frequency bumps. PDAF processing causes periodic ones (eg PDAF banding). You can sometimes see where the stepper exposed the chip nonuniformly. The G1 and G2 channels often have different spectra.

Jim

Agreed. So have I.

As an aside, look at the noise on each pixel in the Raw channels. Most RAW images are in tiff containers and Matlab has a handy tiff reader that breaks down each of the objects so it's possible to examine the pixels for various noise sources. The unavoidable is photon noise which is a Poisson distribution which becomes gaussian as the signal level increases. Then there is thermal noise which is similar but is a relatively fixed level based on the sensor and electronics temperature. Pattern noise is often associated with noise that couples from the memory access logic and step differences in the AtoD converters.

A 2D FFT can give an idea of the relative magnitude of pattern v other sources. AtoD step differences can be identified by making a histogram of pixels. Photon noise can be separated from thermal noise in a model by dark field images at different shutter speeds and adding "almost dark" field shots where the pixel values are one percent or so of full dark field images.