Abstract: Classic statistical thresholding methods take class variance sum of some form as criterions for threshold selection. They don't take special characteristic of practical images into account and fail to get ideal results when segmenting a kind of image having similar statistical distributions in the object and background. In order to eliminate the above limitation of classic statistical approaches, a novel statistical criterion was defined by utilizing standard deviations of two thresholded classes, and the optimal threshold was determined by minimizing it. Experiments on a variety of infrared images and general real world images show that our method outperforms the existing classic thresholding methods in segmentation quality, especially for infrared images.

Key words: Image segmentation, Thresholding, Statistical theory, Standard deviation