Anette Gottfridsson
David Grimfors

Titel:

Demosaicing with simultaneous noise reduction

Abstract:

Digital images are represented as colour triplets, eg. RGB (red, green,
blue), but digital camera sensors usually capture only one colour in each
pixel. In addition, the image is corrupted by noise. The unknown colours
in each pixel are interpolated in a process called demosaicing, sometimes
followed by noise reduction.

It is however also possible to perform the noise reduction before or
simultaneously with the demosaicing, though this is seldom done
today. In this report we have evaluated these possibilities with focus
on algorithms suitable for real time hardware implementation in
digital video cameras.

We will develop a simple yet effective noise reduction method for use
before demosaicing, and a very powerful although fairly complex method
for performing both operations simultaneously using Gaussian Markov
Random Fields (GMRF) where we have used an adaptive neighbourhood
structure and incorporated colour correlations. We have used
overlapping image patches to speed up execution and we will show that
local image variance is a simple and efficient way to estimate model
parameters.

In practice, noise reduction is a tradeoff between leaving noise and
losing detail. We will finish by discussing noise models to help
adjusting this balance and conclude that accurate noise estimations
are important, especially for advanced noise reduction methods.

We will conclude that all three orders of procedure have advantages
and disadvantages, but for the case of built-in noise reduction in
real time video we find the first approach to be most promising.