Binarization of noisy gray-scale character images by thin line modeling

Abstract

In this paper, we propose two new methods for the binarization of noisy gray-scale character images obtained in an industrial setting. These methods are different from other conventional binarization methods in that they are specially designed to detect only character-like regions. They exploit the fact that characters are usually composed of thin lines (strokes) of uniform width. We first model the shape of the cross section of a character stroke and discuss how to detect the character stroke. Then, ALGORITHM I, which is a direct realization of our basic idea, is introduced, followed by an advanced algorithm named ALGORITHM II. The key to these algorithms is the local binarization-voting procedure. The performance of our methods is evaluated and compared with that of five other binarization methods using 550 slab ID number images, where a common character segmentation routine is attached to each of the different binarization routines and the segmentation success rate for each method is obtained. Experimental results show that ALGORITHM II results in far better performance than the other methods.