Thresholding techniques are fundamental for region segmentation of a g
ray level image. It is often realistic to assume that gray pixel level
s are subject to a mixture of normal distributions. This paper points
out problems caused by the underestimation of variances from a frequen
cy distribution and solves the problems by transforming a frequency di
stribution into a piecewise-continuous distribution based on the maxim
um entropy criterion. This transformation improves conventional thresh
olding techniques. Experimental results showed the effectiveness of th
e transformation.