A. A. Zaitseva

Different representation formats exist for efficient video data processing, transmitting and storing. To solve the problem of image restoration after lossy compression the method based on hierarchical representation of two-dimensional data stream and adaptive dynamic segmentation (ADS) is suggested to use in compression stage. Depending on problem the type of structurization and its program realization rely on revelation of parametrical estimations for ε-identity criterion. From the point of view of recursive-fractal approach the semantic components of information object reveals via iterating process of ADS. In image and video processing the structurization is based on exposure of conceptual fragments out of two- or three-dimensional structure of data representation. The method of adaptive dynamic video data representation originates from iteration-function mathematical model (theorem of collage). The common scheme of adaptive dynamic segmentation can be presented as follows: the ε = 0 gives original "pixel-to-pixel" representation of image, further increase of ε value gives «in-layer» representation. Each of the layers reflects ε-identity to original image. The given value of ε carries the qualitative characteristic of original image «desensitization» by joining of pixel values into segments (layers) according to ε level. This algorithm as the theorem of collage has the following stop criterion: semantic search with given maximum ε which doesn-t demand full enumeration with ε → 0. For arbitrary image by the use of the ADS results it is possible to build an ADS tree (semantic tree) in two different ways: «top-down» and «bottom-up» segmentation. In the first case the tree starts to build from its root node (ε takes its maximal value and the whole image consists of one fragment) and than the value of ε iteratively decreases. The benefit of this approach is the simplicity of semantic tree building process. But this simplicity implies possibility of distortions in the results. In the second case the segmentation starts from terminal nodes of the tree (ε takes its minimal value and the quantity of fragments is maximal). This allows simplification of image segmentation process but the difficulties in semantic tree restoration emerge (the root node search