in computer vision, noise is assumed to be normally distributed with a mean of 0 and some standard deviation σ
Scale
size of structures of interest
small scale structures may be noise
when we have small scale structures in an image (or a series of still images), it might be noise or actual structures. if it is noise, it will vary across images.
Signal to Noise Ratio
SNR = maxsignal/ σ noise
higher means (more signal per noise) so better images
Noise Reduction
Temporal Averaging
If we have multiple still images
averaging the images reduces the noise
if we average N images with noise of standard deviation σ, the noise reduces by a factor of sqrt(N)
σ noise = σ / sqrt(N)
Spatial Averaging
a.k.a. Local Averaging
The idea is that the neighbourhood of a pixel is similar to the pixel
replace centre pixel value by average of the neighbourhood pixel value including the centre pixel
The higher the neighbourhood, more the noise goes down BUT the picture becomes more blurry/loss of sharpness/loss of spatial information/ loss of small scale structures
this happens because the principle spatial similarity does not hold true for edges