Robust local scale estimators for blind evaluation of noise variance in images

Abstract

A task of blind evaluation of additive Gaussian noise variance in images is considered. One stage of many methods used for this purpose is obtaining local estimates of noise variance in blocks (scanning windows) of small size tessellating an image. For this purpose, we propose to use robust estimators of sample scale. Several such estimators are studied for cases of pure additive noise and mixed additive and impulsive noise. It is shown that the use of robust scale estimators allows increasing the number of “normal” estimates, close to a true value of noise variance. However, the estimate distributions have specific features to be taken into account in searching distribution mode accepted as the final estimate of noise variance. Efficiency of noise variance estimators based on local robust scale estimates is evaluated for color image database TID2008. It is demonstrated that for the pure additive noise case this efficiency is comparable to that one of standard methods. In turn, for the mixed noise case with probabilities of up to 0.1 in each color component, the proposed approaches are still able to provide appropriate estimation accuracy of variance for additive noise component.