Jump to content

Banner.jpg.b89429c566825f6ab32bcafbada449c9.jpg

How Random is your Random Noise...

narrowbandpaul

By narrowbandpaul
in Imaging - Discussion

Recommended Posts

narrowbandpaul

narrowbandpaul

Posted
Posted

I found this of interest on the QSI site. It regards the read noise of your camera.

The read noise of your camera should be completely random values. Such random noise images have a histogram that is gaussian. Guassian noise can be reduced by combining frames, so it is important to have a very gaussian read noise.

To prepare the read noise frames, take a seies of bias frames (shortest possible exposure), and average them. The more the better, I used 10, but more would be better. This creates a frame that should have a negligible read noise, but contains the offset value...the bias. Now take a single bias frame and subtract off the averaged bias frames. This give only the read noise of the camera.

This frame will be a mess (once you compress the data by lowering the white point to say 100). Although a mess, its exact messiness is important. Remember, if the noise is gaussian, then it will be totally random...no patterns. If you see a pattern emerging, it could be bad.

To properly see how gaussian your read noise is, we need to perform a mathematical operation called a Fast Fourier Transform. This allows you to see the frequency component of the read noise.

To perform this FFT, you can download a free program called ImageJ, which does the FFT in a single click. Very easy to use.

Load in your read noise frame. For the FFT to work properly we need to have a square image whose length is a power of 2, i.e. 256x256, 512x512, 1024x1024. 1024x1024 is more than enough for the FFT. Once this is done click Process, FFT, FFT, and a new image will pop up..the FFT of your read noise.

What should we look for...

there should be a single bright pixel in the centre of the image. But apart from that the image should be totally random, much like your read noise frame. There should be no patterns. if there are pattern then your read noise is not gaussian, and so a part of your read noise cannot be eliminated by stacking. One interesting property of the FFT is its orthogonality...if vertical patterns appear in the FFT, then the problem in your original read noise is horizontal.

In ImageJ, or any other programme that displays a histogram, look at the histogram of your FFT. This should be gaussian too... one property of a fourier transform is that the Fourier Transform of a gaussian is itself gaussian. The nearer the FFT histogram looks to gaussian, the better your read noise.

I have performed this for the H16 and H9 cameras from Starlight. the H16 looks very good, only a slight elongation in the vertical direction. The H9 has some vertical streaks, meaning a non gaussian read noise.

They are displayed below

7573_normal.jpeg

(click to enlarge)

7574_normal.jpeg

(click to enlarge)

So why bother with all this...

if you cant remove all the read noise from your images then you hurt your dynamic range. Dynamic Range=Full Well Depth/read noise. hurting your dynamic range reduces the ability of your camera to pick up faint detail.

Futher more, one can measure the read noise of your camera quite easily. Go to your original read noise frame. You need to take a line profile (intensity of pixel across a line in the image). I use imagesplus, which saves the line profiles as txt files, which excel can read. Perhaps with other programmes you can read an image staright in to eg Matlab.

Load your line profile into excel. We only need the intensities. multiply each intensity by the gain of your camera. Gain=well depth/65535 if you have a 16 bit camera. Raws from DSLR's are normally 12 bit. This provides the number of electrons stored in each pixel. Take the standard deviation of these numbers. Stdev(...) in excel. The number is the number manufactureres quote when they say read noise..."..."

I found this interesting, so thanks for reading..

i ask you, how random is your random noise...

Paul

PS, darks are cool too, as are single stars for FFT analysis

Link to comment
Share on other sites
MartinB

MartinB

Posted
Posted

I've seen that article Paul and have been thinking of giving it a try. Noise is far more complicated than we think and doesn't behave in the way we think it should. I'm still not sure about constant read noise which is random isn't removed by stacking!! Must find out more.

Link to comment
Share on other sites
AlphaTau

AlphaTau

Posted
Posted

So Paul, lets see if I understand this.

If the noise is random you can remove, or more properly, fill in the gaps between the random noise by adding other images with noise in a different location on the image (stacking). But if the pattern is fixed and you add other images you just increase the noise in the location of the fixed pattern ......Is that right?

Very interesting read thanks :thumbright:

Andrew

Link to comment
Share on other sites
narrowbandpaul

narrowbandpaul

Posted
  • Author
Posted

remember its the read noise you need to look at in ImageJ...ie (single bias-averaged bias).

If your pixel values are really random, the averaging a set of these frames smooths out the variation (standard deviation) in pixel value. If there is a pattern present in all frames, the averaging only reinforces this signal. Subtraction will remove it, but averaging will not. Which is what you said?

Glad to see you had a go psychobilly. i am interested in the read noise rom the DSLR's. i have a 300D but havent taken any frames yet.

Cheers Guys

Paul

Link to comment
Share on other sites
Psychobilly

Psychobilly

Posted
Posted

psychobilly, why dont you post your FFT's on here so we can see...could form a basis of a comparison of different DSLR and CCD chips...

I'll take some frames myself (in RAW) and see what I get

I think i did it right...

I used DSS to stack 9 bias frames with average selected as the mode - also had to let it process a single light frame to get it to do anything...

I copied the master offset file to "Average9bias"

I opened this in cs3

Opened a single bias frame

Selected the whole bias frame and pasted into the average image

Selected difference mode for the layer op and flattened layers

Adjusted white point to 100

set crop to 20.48cmx20.48cm at 100 picxels per cm and cropped a 2048x2048 out of the image.saved this as Average-single2048x2048

Processed this crop in ImageJ

Save the result as a tiff and resized it to 512x512 and jpeged it

had 4 crosses equidistant from each corner - bit liek teh diffraction pattern from a spider vane?

Looked at the histogram and it looked pretty gausian to me...

Was this the right way to do things?

Billy...

Link to comment
Share on other sites
narrowbandpaul

narrowbandpaul

Posted
  • Author
Posted

i dont use photoshop but as long as you subtracted the single frame from the averaged bias, and then cropped to a square image of length 2^n, then thats fine. it sounds like you did this...

did you see the diffraction pattern in the FFT of a star, or was your read noise FFT like a sipder vane diffraction? Is so, it doesnt sound to great. t sounds like it looks like the FFT of a sine wave pixel intensity variation. Perhaps your noise is composed of different 'frequencies'.

Glad your histogram looked gaussian

Paul

Link to comment
Share on other sites
narrowbandpaul

narrowbandpaul

Posted
  • Author
Posted

i have now done the analysis of my Canon 300D at ISO 800 at room temperature. I took 20 bias frames (1/4000th exposure) and averaged them. The results arent great im afraid.

There was a noticeable pattern of vertical and horizontal stripes in the averaged bias frames...pattern noise being reinforced. The read noise itself looked not to bad, but it was noisy.

On taking the FFT of the read noise, some strange things were observed.

There is an elongation in the central column of pixels..similar to the H16 but a bit more obvious. The more curious thing is the distribution of pixel intensities in the FFT image. There is a central hotspot, like seen in vignetting. This hotspot was not observed in the original frames. I dont know whats causing this but it is very strange. The histogram has a sinificant deviation from gaussian.

On taking the FFT of the averaged bias...

A bright line is seen across the entire central row...caused by the patterning i saw earlier. Again the vignetting is present.

Here are the images...FFT of the read noise first

7617_normal.jpeg

(click to enlarge)

7618_normal.jpeg

(click to enlarge)

Both images showed significant deviations from a gaussian histogram.

I also looked at the standard deviation of the pixel intensities acroos one row, of the read noise image, in excel.

The standard deviation at room temp and ISO 800 is 47.47 intensity units at 12bits

the standard deviation of the 20 averaged frames fell to 14.786, 3.2x lower. The average value of the averaged frames approximates well the offset. this value was measured at 102 units of intensity.

Does anyone know the gain of the canon camera for 12 bit raws at ISO800...?

Best Wishes

Paul

Link to comment
Share on other sites
MartinB

MartinB

Posted
Posted

it probably can be removed, by subtraction, but not by stacking...the fixed pattern just reinforces itself

If this were true then there is very little point in taking long exposures. The whole point of long exposures is to make read noise an insignificant proportion of the overall noise. Bias and fixed pattern noise is easy to deal with by subtraction. Theres more to this that I don't understand.

Link to comment
Share on other sites

Archived

This topic is now archived and is closed to further replies.

Go to topic listing

  • Recently Browsing   0 members

    • No registered users viewing this page.
×
×
  • Create New...

Important Information

We have placed cookies on your device to help make this website better. You can adjust your cookie settings, otherwise we'll assume you're okay to continue. By using this site, you agree to our Terms of Use.