A generalization of Otsu method for linear separation of two unbalanced classes in document image binarization

The classical Otsu method is a common tool in document image binarization. Often, two classes, text and background, are imbalanced, which means that the assumption of not met. In this work, we considered imbalanced pixel classes background text: weights different, but variances same. We experimentally demonstrated employment criterion takes into account imbalance classes' weights, allows attaining higher binarization accuracy. described generalization criteria for two-parametric model, an algorithm optimal linear separation search via fast clustering was proposed. also model with proposed increasing accuracy documents complex or spots.