IMPROVEMENT OF THE IMAGE CLUSTERIZATION METHOD

Abstract

Image clustering algorithms and their possible improvement are considered in the paper. Most image segmentation algorithms focus on a certain limited class of images, which leads to a reduction in the scope of their application. The problems of the existing algorithms include the automation of the process of dividing the image into domain-blocks, the slowdown of compression, the limitation of the image size, and the complexity of the algorithm. This is due to the definition of the minimum number of images that have common features with the partition standard. The direct impact of noise, the coincidence of elements in the structure of class representatives, lead to a significant increase in the image coverage area, which leads to a decrease in the accuracy of clustering. Image clusters are formed based on criteria such as color, texture, and shape. The use of the fractal dimension of the image as a criterion for clustering and for setting the number of partition blocks allows to expand the range of application of the algorithm and increase the speed of data processing without losing the accuracy of the algorithms. The given algorithm for image segmentation is not tied to a specific image and does not require complete similarity with the standard. As a form, it is proposed to use a deterministic fractal obtained by fractal approximation of the image.
Such an algorithm also has its drawbacks, which include the limited number of groups of fractals that have already been studied. However, the rapid development of fractal geometry methods will make it possible to get rid of this drawback. The operation of the algorithm was tested for fractal image compression, for dividing the image into domain blocks.

The proposed algorithm allows you to use the given improvement not only for image clustering, but also for any object clustering and fractal image compression.
Key words: clustering, image, fractal approximation, domain blocks, fractal dimension, deterministic fractal.