INFORMATION-EXTREME ALGORITHM OF THE SYSTEM FOR RECOGNITION OF OBJECTS ON THE TERRAIN WITH OPTIMIZATION PARAMETER FEATURE EXTRACTOR

Abstract

Context of this article topics is that the issue of the choice of parameters of the extractor features and classification analysis algorithm in conditions of a priori uncertainty, resource constraints, and information is not enough studied and in full has not been decided so far.Objective: to increase efficiency of functioning autonomous system of recognition in information and cost sense which functions in the modes of training and examination in the conditions of limited volumes of the training dataset by optimization of parameters feature extractor and classifiers.Methods of a research are based on algorithms of digital processing and analysis of images for formation descriptors of interest objects, on the principles of mathematical statistics, the theory of information for assessment of functional efficiency of decision rules, provisions on the theory based on population search algorithms for optimization parameters of scanning images system.Results: the developed machine learning algorithm with rough observations binary coding and modification swarm optimization algorithm recognition system operating parameters allow to obtain for small volume dataset decision rules with reliability which approaches boundary maximum value. This experiment shows the advantage of the use swarm algorithm for scanning images, which is three-fold increase in performance compared to known algorithms RASW and ESS.Conclusions. Proposed method for the synthesis of information extreme classifier of images with rough binary encoding of sparse histogram of frequency of occurrence of visual words, to provide a computationally efficient decision rules. Improved method based on population search to adjust parameters of the extractor features that allows you to get the best value in the information and cost meaning ofthe parameters functioning system recognition of images in a few iterations of the algorithm work. The practical value of the results is to obtain well-functioning designing algorithms capable of learning image recognition, which operates under conditions of resource limitations and information.