Ensembles of normalization techniques to improve the accuracy of Otsu method

Abstract

Otsu method is a global thresholding method that uses between class variance as a discriminant criterion in order to maximize the separation between background and foreground of an image. However, there are problems of biasness in Otsu method. These problems are caused by the differences in class variances. The threshold value obtained by Otsu method will be bias towards the larger class variance component. Hence, in this paper, a new variant of Otsu method by using normalization techniques and their ensembles is proposed. By using normalization techniques, grey level values will be transformed into a smaller range in feature space and this will affect Otsu method as this method depends on grey level values. The domination of certain grey level values also will be eliminated. Rank filtering 1566 Fauziah Kasmin et al. has been applied to eliminate noises and ensemble approaches of normalization techniques are utilized to increase the performance of the proposed method. Ensemble approaches namely Maximum Variance, Majority Voting, Product Rule, Addition Rule and Average Rule have been applied on the binary images obtained. From the experiment on 20 retinal images randomly selected in 50 runs from DRIVE and STARE databases, Maximum Variance shows the most significant result that is 95.39% accuracy. While from the experiment on 15 document images randomly selected in 50 runs from DIBCO2009 and DIBCO2011 databases, Average Rule shows the most significant result that is 97.17% accuracy. This indicate the use of ensembles of normalization techniques can give promising result to improve Otsu method.