Sensor noise is a random variation in pixel values, but what kind of random is it? Does it follow a normal distribution, one of the gamma distributions, or something funky like an F distribution?

What applications does this knowledge have?


1 Answer 1


There are many different sources of noise in images with different distributions. For example shot noise, which is a large contributor in low light arises from the random emission of photons follows the poisson distribution. Dark current noise and read noise (major contributions in shadow noise in good light) are more complex as they exhibit banding and are rarely evenly distributed.

Regarding applications of this knowledge, most commercial noise reduction software works by fitting a distribution to the image data by finding an area of constant colour (either automatically or by user intervention) to estimate the parameters of the distribution so knowing these values in advance is not important.

  • 2
    \$\begingroup\$ Could you expand on Dark current and read noise? \$\endgroup\$ Apr 25, 2011 at 21:12
  • \$\begingroup\$ Which noise reduction algorithms are you talking about? I know that Non-Local Means (Buades, 2004), Total Variation (ROF, 98), and anisotropic diffusion (perona and malik) don't. Or are there some commercial packages that don't have published algorithm's that you're talking about? \$\endgroup\$
    – mmr
    Apr 25, 2011 at 21:40
  • \$\begingroup\$ @mmr Back when I used the commercial software Neat Image, it looked for a detail-free patch to use for noise estimation, and you could "aid" it by selecting such a patch manually. \$\endgroup\$
    – coneslayer
    Apr 25, 2011 at 22:14
  • \$\begingroup\$ @mmr Yes I was talking about noise reduction software rather than published algorithms, edited to avoid confusion \$\endgroup\$
    – Matt Grum
    Apr 26, 2011 at 14:09
  • \$\begingroup\$ Dark current is the current generated in a sensor due to what happens when a photon hits a pixel happening instead at random without a photon. Shot noise is due to photons being discrete units of light and when you're down at a certain low level of light you're relying on "luck" which will play out in your favour in some pixels and not others. (Think of it like roulette: every cell bets on 23 (receiving a photon). 1 in 38 (is it?) pixels will win during your shot, the rest lose. You get noise. With enough photons/time, the real pattern becomes obvious because you average out. \$\endgroup\$
    – Dannie
    Sep 18, 2019 at 21:34

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

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