Sakalo I. Frame image processing based on artificial neural networks

The object is a process of frame imaging warped by different types of interference under the conditions when data for processing arrives in a successive order in the online mode. The goal is to develop high-speed frame imaging techniques based on specialized artificial neural networks, which are self-learning. Methods are the basic points of the image processing theory (for stating the problem being solved), artificial neural networks (for synthesis of specialized architectures), optimization and identification (for synthesis of high-speed learning methods of the proposed neural networks), linear algebra and mathematical statistics (for studying the properties of the synthesized methods and architectures). Apparatus - personal computer. Theoretical and practical results - developed neuron networks and methods of their self-learning can serve as basis for the construction systems of treatment video information in the online mode in different spheres of medicine, industry, medias applications. Scientific novelty lies in the unique development of a competitive neural network and a method of its self-learning where fragments of images are used as input signals instead of vector signals, which provides consideration of inter-pixel correlation and differs in computational simplicity, has filtering properties and is capable of processing non-stationary video signals in real-time; unique development of a specialized neural network for simultaneous analysis of principal and independent components and the learning method in real discrete time with a high-speed performance, which provides an optimum choice of parameters (in the sense of reducing the average error) and allows simultaneous solution of both the problem of image compression, and the problem of blind signal identification and separation; improvement of learning methods of neural networks for solving the tasks of image compression, which have increased performance due to prior optimizing of their synaptic weights and allow to process data in real time as it becomes available; improvement of self-learning methods for T. Kohonen's neural networks on the basis of the previous signal filtering and use of robust criteria, which allows to solve the problem of segmentation of images disturbed by intensive noise. The results of the thesis were used while making a system for managing tube furnaces and an identification subsystem in Pobuzhsky Ferronickel Plant LLC (act of implementation from 22.09.2010); while using a camera of panoramic view to identify images in the processing of video recordings of football matches in JSC "Metalist" (act of implementation from 14.09.2010) and in preparation of bachelors, specialists and masters in Computer Science at Kharkiv National University of Radio Electronics (act of implementation from 17.09.2010). Scientific and practical results of the thesis can be used: in computer vision systems, bio-medical research using visual information and media communications for high-quality digital image processing provided that information is received for processing in the online mode.