1
$\begingroup$

I'm trying to add non-Gaussian noise to an image. The purpose is to synthesize realistic camera noise. I am doing the following:

  1. Obtain a noisy image sample by subtracting a low ISO photo from a high ISO of the same photo.

  2. Find a flat 128x128 section of image with no edges by randomly cropping and checking if the standard deviation is small enough.

  3. Zero mean the patch.

  4. Normalise the patch by the patch standard deviation.

  5. Compute PSD (power spectral density) by windowing and then using fast Fourier transform (fftn in Pytorch). Followed by squaring the frequency domain representation.

  6. Generate Gaussian noise of same standard deviation as cropped noise patch.

  7. Compute frequency representation and PSD of noise as above.

  8. Do the following:

    H = gaussian_noise_psd / image_psd 
    output = image_frequency_domain * H
    
  9. Lastly I convert back to spatial domain by reverse FFT and reversing the normalization step.

My output doesn't look quite right. The generated noise seems to be a pattern and doesn't look very natural. I'll put a sample below. I've added 0.5 to the output as the noise is zero mean.

Cropped area from noisy sample

Noise crop in original image

Generated Gaussian noise of same std as above

Generated Gaussian noise of same std

Output of experiment

Output of my experiment

While the first two images have a std of 2 in this case, the output image has a much higher std of 7.5. You may need to open the images in full screen as they are quite small. How do I replicate the noise properly?

$\endgroup$
3
  • $\begingroup$ Why you believe gaussian noise is not a good model? Spatial dependency? are you able to capture images? if so, why not take several photos of a wall and reconstruct the noise distribution? What do you need this noise model for? $\endgroup$
    – Bob
    Sep 13 at 16:37
  • $\begingroup$ I actually have been working with Gaussian noise up to now but I'm trying to explore the efficacy of neural network models that have been trained with Gaussian noise on 'real world images'. In other words, machine learning models are trained with Gaussian noise but this isn't how camera sensors apply noise. Id like to extend this to training with more realistic noise that can be synthesised. Anyway I think I might have to go back to the drawing board! $\endgroup$ Sep 13 at 17:05
  • $\begingroup$ “Followed by squaring the frequency domain representation.” Why? If you want to give the spectrum of the Gaussian noise the same shape as the spectrum of the noise image, don’t square. Just multiply the Gaussian noise by the magnitude of your noise image in the frequency domain. $\endgroup$ Sep 14 at 0:07
1
$\begingroup$

You can create noise with the same frequency-domain characteristic as your example noise image simply by multiplying a white noise frequency spectrum by the magnitude of the frequency spectrum you are trying to emulate. White noise is expected to have equal power at all frequencies (though this will not be exactly the case), so the multiplication will "copy the shape".

Setting the standard deviation of the noise can be done after the fact.

This is example code in MATLAB using DIPimage (disclosure: I'm an author). But I think it's simple enough that you can consider it pseudo-code and replicate it in any other language with any other image-processing library.

a = readim('https://i.stack.imgur.com/JaeTF.png');
a = a - mean(a);                     % must be zero mean
b = noise(newim(a),'gaussian', 1);   % create white Gaussian noise image
c = real(ift(abs(ft(a)) * ft(b)));   % multiply in frequency domain
c = c * (std(a) /  std(c));          % scale to get the right std. dev.

Here are images a, b and c, stretched for display:

a b c


PS: Note that there's a difference between "Gaussian noise" (noise created by sampling a Gaussian distribution) and "white noise" (where each value is uncorrelated to other values). You can do the above with other distributions too, try starting with uniform distributed noise instead of Gaussian, for example. b is white noise, c is not. But c is still Gaussian!

$\endgroup$
2
  • $\begingroup$ Thanks, you've given me a lot to work with! My samples make a lot more sense now. One last small question: Should I drop the window function prior to ft()? I seem to be getting a slight circular brightness gradient in my output when using the window function. $\endgroup$ Sep 14 at 14:44
  • $\begingroup$ @BledClement Windowing is important in some estimation tasks, but I don't think it's needed here. Note that you would be able to tile your noise patch and not see the seam very clearly. The window function serves to remove that seam in cases it is obvious. If your results are worse with a window function, you should definitely leave it out. $\endgroup$ Sep 14 at 16:10

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.