Improving the Effectiveness of Image Classification Structural Methods by Compressing the Description According to the Information Content Criterion.

The research aims to improve the performance of image recognition methods based on a description in the form of a set of keypoint descriptors. The main focus is on increasing the speed of establishing the relevance of object and etalon descriptions while maintaining the required level of classification efficiency. The class to be recognized is represented by an infinite set of images obtained from the etalon by applying arbitrary geometric transformations. It is proposed to reduce the descriptions for the etalon database by selecting the most significant descriptor components according to the information content criterion. The informativeness of an etalon descriptor is estimated by the difference of the closest distances to its own and other descriptions. The developed method determines the relevance of the full description of the recognized object with the reduced description of the etalons. Several practical models of the classifier with different options for establishing the correspondence between object descriptors and etalons are considered. The results of the experimental modeling of the proposed methods for a database including images of museum jewelry are presented. The test sample is formed as a set of images from the etalon database and out of the database with the application of geometric transformations of scale and rotation in the field of view. The practical problems of determining the threshold for the number of votes, based on which a classification decision is made, have been researched. Modeling has revealed the practical possibility of tenfold reducing descriptions with full preservation of classification accuracy. Reducing the descriptions by twenty times in the experiment leads to slightly decreased accuracy. The speed of the analysis increases in proportion to the degree of reduction. The use of reduction by the informativeness criterion confirmed the possibility of obtaining the most significant subset of features for classification, which guarantees a decent level of accuracy. [ABSTRACT FROM AUTHOR]