Abstract
This paper describes a study on improving Otsu method by using normalization techniques and their ensembles. Otsu method is known as a global thresholding method that use discriminant criterion, between class variance, to maximize the separability between background and foreground. However, Otsu method fails to threshold unimodal images. Variance is easily affected by changes of intensity values. Due to that factor, normalization techniques have been used in this study where two normalization techniques have been applied on a particular input image at one time. First, column vector is transformed into zero to one as feature vector is in the form of column vector. Then, another four normalization techniques namely L1-norm, L1-sqrt, L2-norm and L2-hys have been applied on the image consecutively. Ensemble approaches of these normalization techniques have been proposed to increase the performance of Otsu method. Maximum variance, majority voting, product rule, addition rule and average rule have been applied on the binary images obtained. From the experiment on 50 images, product rule shows the most significant results.