LABORATORY 13
Laboratory of Systems for Behavior Organizing
Head of Laboratory – Dr. Modest Vaintsvaig
Tel.: (095) 209-42-25; E-mail: wainzwei@iitp.ru
The leading researchers of the laboratory include:
Dr.Sc. (Techn.) |
V. Neiman |
Dr.Sc. (Techn.) |
A. Tsybakov |
Dr.Sc. (Math.) |
P. Nickolayev |
Dr. |
A. Shen |
Directions of activity:
MAIN RESULTS
Using the tree-structured procedure for matching 2D-images we have developed an algorithm for point-wise matching of 2.5D-images that distinguishes matching and non-matching regions. We have developed a general scheme for a procedure that looks on moving 2D and 2.5D images and finds boundary points for surfaces generating those images and their trajectories. We constructed an initial version of a cluster-finding algorithm that generates "visual notions" later used for recognition.
We developed a novel method of image segmentation (the task of boundary type identification for different color regions with highlights, shadow, etc.) that allows to use Gaussian, zonal and linear approximation models of base spectral functions. This method allows to get an invariant three-parametric estimate of reflectance function of surfaces for scene image made using many light illuminants having different chromaticity. In the framework of linear theory of spectral stimulus formation we have developed four algorithms for robust clusterization of 3D-color space using Generalized Hough Transform. These algorithms can be applied to scene images of different complexity.
We developed a criterion for hypotheses testing for binary images where parametric default hypothesis is tested against a non-parametric alternative hypothesis. It is shown that this criterion converges as fast as possible. We get an exact asymptotic bound for a minimax adaptive risk for the problem ofestimations linear functionals in n-dimensional case. We have found an adaptive estimate that reach that bound. We have constructed an optimal prediction algorithm for the linear regression model with infinitely many parameters (assuming that parameters are limited by a given function that tends to zero as index goes to infinity). This method allows to motivate the elimination procedure for non-informative parameters.
We have investigate connections between combinatorial and algorithmic definitions of the information (in Kolmogorov's sense). We found a natural combinatorial interpretation for linear inequalities that include Kolmogorov complexities and have only one term in the left-hand side. We proved that each inequality of this type is a positive linear combination of the so-called basic inequalities.
We have investigated the connection between degree functions, chaotic mappings, fractals and statistics for process with long-term dependencies. We have suggested a simple model for computing network resources in the case of self-similar teletraffic.
GRANTS FROM:
PUBLICATIONS IN 1999