Bobniev R. Neural network methods and means of image compression

The thesis covers the analysis of the problem of image compression using artificial neural networks (ANN). The most frequently used static ANN architectures and learning algorithms are investigated for their application in solving the classification, clusterization, vector quantization, approximation and image compression problems. Several methods of methods and architecture modification are proposed with the purpose of elimination problems related to ANN work. The problem of input data normalization and ability to use weights of the network, have been solved by adding "normalization" input. The hybrid genetic algorithm (GA) is proposed, which is modification of the classical GA utilizing the feature of biological apoptosis. The main difference is taking into account number of the identical individuals in population and killing them at the crossing stage or natural selection stage. A method of initial input data clusterization and classification is proposed to initialize centers and widths of basis functions for ANN radial-basis functions. The determination of centers is made by means of Kohonen's and Neural-Gas networks. The determination of widths is made by means of k-means and k-neighbours algorithms. Simulation in the Microsoft Visual Studio 2010 Express Edition environment shows the high efficiency of using ANNs for solving various problems of approximation, classification, clusterization, vector quantization and image compression.