An adaptive median based filter is proposed for removing noise from images.
Specifically, the observed sample vector at each pixel location is classif
ied into one of M mutually exclusive partitions, each of which has a partic
ular filtering operation. The observation signal space is partitioned based
on the differences defined between the current pixel value and the outputs
of CWM (center weighted median) biters with variable center weights. The e
stimate at each location is formed as a linear combination of the outputs o
f those CWM filters and the current pixel value. To control the dynamic ran
ge of filter outputs, a location-invariance constraint is imposed upon each
weighting vector. The weights are optimized using the constrained LMS (lea
st mean square) algorithm, Recursive implementation of the new filter is th
en addressed. The new technique consistently outperforms other median based
filters in suppressing both random-valued and fixed-valued impulses, and i
t also works satisfactorily in reducing Gaussian noise as well as mixed Gau
ssian and impulse noise.