| 摘要 |
FIELD: physics, video.SUBSTANCE: invention relates to image binarisation means. The method includes selecting a number of points from a reference image, calculating predetermined factors for said points, selecting from said factors those factors in whose space distinct clusters exist, said clusters being formed by points of an object and a background; determining in the found space training samples for classifiers consisting of medoid points of each cluster and boundary points which divide clusters. In the method, said samples are used during operation of two classifiers of the "K nearest neighbours" type, to the input of which all points of the processed image are successively transmitted, where the method comprises using first a classifier based on boundary points in a "qualified majority" version, then for points not classified by the first classifier, using a classifier based on medoid points in a "qualified majority" version, and for the remaining points a classifier based on boundary points in a "simple majority" version; adjustment of the classifiers is completed by selecting the training samples.EFFECT: high quality of binarisation.6 dwg |
