V International Conference on "Information Technology and Nanotechnology" (ITNT-2019) An investigation of machine learning method based on fractal compression E Y Minaev 1,2 1 Samara National Research University, Moskovskoe Shosse, 34А, Samara, Russia, 443086 2 Image Processing Systems Institute of RAS - Branch of the FSRC "Crystallography and Photonics" RAS, Molodogvardejskaya street 151, Samara, Russia, 443001 e-mail: eminaev@gmail.com Abstract. In this article the method of machine learning with cyclic fractal coding and the use of domain block dictionary, adapted for use on mobile platforms, with optimization of performance and volume of stored fractal images is investigated. The main idea of the method is to use the fractal compression method based on iterated function systems to reduce the dimension of the original images, and to use cyclic fractal coding to represent the class of images. As a result of research of the method it was found that the share of correctly recognized objects on MSTAR averages 0.892, the recognition time averages 254 ms. The achieved results are acceptable for use in mobile platforms, including UAVs and ground autonomous robots. 1. Introduction The problem of using existing fractal compression algorithms on mobile hardware and software platforms is noted in [1]. Traditionally, fractal compression methods have high computational complexity, and methods and algorithms for optimizing performance developed for desktop hardware platforms are not always applicable for mobile platforms [2] [3]. Modern performance solutions are based on the use of user-programmable gate arrays (FPGAs) and the use of GPUs, which makes it difficult to use these approaches for most mobile platforms. At the same time, the urgency of using fractal compression methods for mobile devices is emphasized in the article [4]. 2. Implementation of machine learning method based on cyclic fractal compression One of the promising approaches to the implementation of the classifier based on fractal compression is proposed in [5]. When we trained the classifier described in [6], the main problem was that the images forming the training sample of one class were compressed independently of each other, and were combined together only at the stage of construction of the support subspaces. At the same time, the recognition stage raises problems associated with the possible intersection of the support subspaces. Accordingly, it is necessary to apply methods that provide spatial separability, which further increases the computational complexity. In [7], a fractal compression scheme using several