Two modified Otsu image segmentation methods based on Lognormal and Gamma distribution models

Abstract

Otsu's method of image segmentation is one of the best methods for threshold selection. With Otsu's method an optimum threshold is found by maximizing the between-class variance; Otsu algorithm is based on the gray-level histogram which is estimated by a sum of Gaussian distributions. In some type of images, image data does not best fit in a Gaussian distribution model. The objective of this study is to develop and compare two modified versions of Otsu method, one is based on Lognormal distribution (Otsu-Lognormal), while the other is based on Gamma distribution (Otsu-Gamma); the maximum between-cluster variance is modified based on each model. The two proposed methods were applied on several images and promising experimental results were obtained. Evaluation of the resulting segmented images shows that both Otsu-Gamma method and Otsu-Lognormal yield better estimation of the optimal threshold than does the original Otsu method with Gaussian distribution (Otsu).