V.M. Khachumov, V.P. Fralenko

The work is devoted to research and development data streams filtering and compression methods that typical for command-and-measurement purpose systems based on a using of ANN and high-performance computing facilities. For these methods are evaluating the effectiveness of the parallel implementation. The filtration problem is solved on the basis of a two-layer feedforward ANN - recirculation network, which is an analog Hopfield network. The peculiarity of the proposed approach in a using of set independent networks, each of which represents a combination of encoding and decoding layers. This approach allows a parallel images processing by parts and reduces learning time. In another experiment the neural network was trained for transformation of noisy images to the prescribed by the expert etalons, that is the classification problem is solved. Also the image compression problem based on the Kohonen network are considered. Images of the compressible flow in the training process of the neural network was divided into disjoint blocks, served in random order to the input network. To increase the compression ratio additionally used an adaptive algorithm to reduce the number of neurons and the zlib library, which use a combination of the LZ77 algorithm and Huffman algorithm, which allowed to prevailing over the algorithm Jpeg in a wide range of compression ratios.