ОЦЕНКА ИНФОРМАТИВНОСТИ И ОТБОР ПРИЗНАКОВ ПРИ ИДЕНТИФИКАЦИИ ОБЪЕКТОВ НА ЦИФРОВЫХ ИЗОБРАЖЕНИЯХ МИКРОСКОПИЧЕСКИХ ПРЕПАРАТОВ, ОКРАШЕННЫХ ПО МЕТОДУ ЦИЛЯ-НИЛЬСЕНА

Autor: Artem Nikolaevich Narkevich, Mikhaylovna Nataliya Koretskaya, Ekaterina Olegovna Zhurbenko, Alina Vladimirovna Kataeva, Konstantin Anatol’evich Vinogradov
Rok vydání: 2017
Předmět:
Zdroj: V Mire Nauchnykh Otkrytii. 9:106
ISSN: 2307-9428
2072-0831
DOI: 10.12731/wsd-2017-4-106-121
Popis: Background. Selecting features for classifying objects in digital images of microscopic preparations, stained by the method of Ziehl-Nielsen with the use of methods for evaluating the informativeness of the Shannon and Kullback, and a comparison of the results of their application. Materials and methods. Used data on 343 687 the selected objects on digital images of microscopic preparations: 6 708 objects acid-fast bacilli, 336 979 – other objects. The analysis of items was carried out for 240 color and morphometric characteristics, evaluating the information content of which was carried out using the methods of Shannon and Kullback. Quality evaluation of selection traits were carried out using the naive Bayesian classifier. Result. The most informative method Shannon had the color characteristics of objects and the ratio of their sizes, and the highest proportion of correctly classified objects (84,6%) were achieved by inclusion in the classification model 6 characteristics with the highest informativeness according to the method of Shannon. The most informative method of the Kullback had radial sizes of objects and their relationships, and the highest proportion of correctly classified objects (78,0%) was achieved when included in a classification model of the 14 signs from the most informative method of Kullback. Conclusion. Methods of estimating the information content of the Shannon and Kullback, can be used in the reduced feature space for classification problems use the method of Shannon allows to a greater extent to reduce the number of signs and thus provides the greatest proportion of correct object classification.
Databáze: OpenAIRE
Externí odkaz: