Threshold selection from image histograms with skewed components based on maximum-likelihood estimation of skew-normal and log-concave distributions
Jinghao Xue and Mike Titterington
Computer Vision and Image Understanding
The purpose of this paper is to propose two new methods for histogram-based image thresholding: one is based on parametric maximum-likelihood estimation of skew-normal distributions, and the other is based on nonparametric maximum-likelihood estimation of log-concave distributions. The main advantages of using these two methods are threefold. First, both methods are natural generalisations of the classical Gaussian-based methods. Secondly, both methods can take into consideration the skewness of the distributions of individual classes. Thirdly, both methods are in line with Otsu's method and the minimum error thresholding (MET) method, based on the comparison of maximum log-likelihoods, for determining an optimal threshold. Compared with Otsu's method and the MET method, the two methods demonstrate comparable, if not better, performance for binarisation of real images and simulated data presented in this paper. Limitations and extensions of the methods are briefly discussed.