Image Description Compression in Classification Structural Methods

Abstract

The problem solved in the article is reduction of computational costs for the image classification process when applying structural methods. The main focus is implementing tools for granulation, screening, and clustering processing a set of elements of etalon descriptions. As a result of compression, each etalon is transformed into a reduced set of descriptors or data centroids, ensuring high speed and performance of image classification. Several variants of simple data compression schemes are assessed and compared to the traditional linear search method, along with two variants of etalon clustering. The comparison includes results achieved for the entire data set and for each of the images separately. The paper presents the results of software modeling of the proposed approaches for two experimental sets containing images of football club logos and artistic paintings. The test sample includes a set of images from the etalon database along with other images that do not belong to the database and with a set of geometric transformations of shift, scale, and rotation in the field of view applied to them. The research covers practical issues of choosing threshold parameters to set the equivalence of descriptors and minimizing the number of class votes to ensure the required level of classification accuracy. Testing has confirmed a significant processing acceleration and a sufficiently increasing level of classification accuracy due to employing compression. Particularly the conducted modeling revealed a tenfold increase in speed. It has been experimentally confirmed that using a clustering apparatus has a much higher potential in terms of classification accuracy and speed than simple sifting or granulation schemes based on close description components.