Improvement of the color text image binarization method using the minimum-distance classifier

Abstract

Optical character recognition systems for the images are used to convert books and documents into electronic form, to automate
accounting systems in business, when recognizing markers using augmented reality technologies and etс. The quality of optical
character recognition, provided that binarization is applied, is largely determined by the quality of separation of the foreground pixels
from the background. Methods of text image binarization are analyzed and insufficient quality of binarization is noted. As a way of
research the minimum-distance classifier for the improvement of the existing method of binarization of color text images is used. To
improve the quality of the binarization of color text images, it is advisable to divide image pixels into two classes, “Foreground” and
“Background”, to use classification methods instead of heuristic threshold selection, namely, a minimum-distance classifier. To
reduce the amount of processed information before applying the classifier, it is advisable to select blocks of pixels for subsequent
processing. This was done by analyzing the connected components on the original image. An improved method of the color text
image binarization with the use of analysis of connected components and minimum-distance classifier has been elaborated. The
research of the elaborated method showed that it is better than existing binarization methods in terms of robustness of binarization,
but worse in terms of the error of the determining the boundaries of objects. Among the recognition errors, the pixels of images from
the class labeled “Foreground” were more often mistaken for the class labeled “Background”. The proposed method of binarization
with the uniqueness of class prototypes is recommended to be used in problems of the processing of color images of the printed text,
for which the error in determining the boundaries of characters as a result of binarization is compensated by the thickness of the
letters. With a multiplicity of class prototypes, the proposed binarization method is recommended to be used in problems of
processing color images of handwritten text, if high performance is not required. The improved binarization method has shown its
efficiency in cases of slow changes in the color and illumination of the text and background, however, abrupt changes in color and
illumination, as well as a textured background, do not allowing the binarization quality required for practical problems