Statistical Methods for Analyzing and Processing Data Components When Recognizing Visual Objects in the Space of Key Point Descriptors

In this paper, we propose the improvement of structural pattern recognition techniques in computer vision systems. We performed the transformation of the key point descriptors space into the space of the data statistical distributions to increase the speed of data processing. These distributions are based on the set of component values obtained by partitioning descriptors into non-intersecting fragments. The compression of data into the fixed set of bits allows to simplify processing and to reduce the quantity of computation operations. BRISK or ORB are suggested to be used as key point detectors, because they form binary descriptors which greatly simplifies processing and analysis. A number of traditional and up-to-date statistical approaches have been proposed and analyzed to determine the relevance of object descriptions by their distribution values and according to the significance level: the chi-square criterion, Renyi divergence, the non-parametric criterion and z-criterion. Additionally, the usage of Manhattan distance between the values of the distribution matrices was tested. The experimental part of the investigations presents the results of calculations and software modeling of the proposed methods for the icons dataset. It is shown, that the statistical distribution models are able to separate set of features effectively even in the case of a small number of bits in the data fragment. An improvement of distinguishing level is confirmed when increasing the size of the fragment in the description structure. The implementation of statistical distributions reduced processing time by hundreds of times preserving sufficient recognition quality.