A new median-based filtering algorithm—extremum median filtering is presented. In order not to perturb the efficient signals as much as possible when the noises are removed

the following approaches are developed in this paper. First

all the pixels are separated into signal pixels and noise pixels according to the decision criterion given in the following; then

noise pixels are replaced with the median value of their neighborhood in the input image. The decision criterion: if a pixel value is the extremum (max or min) of its neighborhood

it is a noise pixel; else

it is a signal pixel. This decision criterion is under such an assumption: inherent relationships exist among neighbor pixels. If a pixel value is far higher or lower than the others' value of its neighborhood are

that is to say

a pixel has lower correlation with its neighbors

we may consider that it had\nbeen contaminated with noise. Else

if it is similar to the others

we consider that it represents an effective signal. Experimental results show that the assumption fits the facts quit well.In this paper

attention is forcused on filtering of images degraded by“salt and pepper”noises. Examples on images containing 184×148 pixels are given.Experimental results show that the EM filtering has better performance than standard median filtering with less subtle details being eliminated. The SNR of the image filtered with EM filter is about 4dB higher than that with median filter. This is because the operation only affects noise pixels and most of the uncontaminated pixels keep intact. Especially

in the case of lower SNR

larger filtering window improves the SNR notably. Median filter is not the case

for the filtering operation blurs the image extremely with the increasing of the filtering window.