Zlotnikov K. A., Dutin G. I., Mishin S. A., Shulika K. M.

In a case when various images of an interest zone with recognizable patterns of objects are statistically disconnected, their complex recognition is decisionmaking on a class of objects as a result of independent identification on separate patterns. In an alternative case, recognition is carried out by the joint analysis of all available set of various patterns. The use of such connections can give considerable effect. In the first variant, at processing of each initial image the recognition device works by a universal principle of a maximum of a posteriori probability. In the article, the expressions describing corresponding solving rules and indicator functions are presented. The funal decision on an object class is accepted either by a principle of the greatest a posteriori probability on the basis of all complex of private decisions, of by a "voting" method. The block diagram illustrating model of complex recognition without an interconnection between patterns is presented. in the second variant, the reduced description of the objects patterns in the form of summary set of the features vectors received on the basis of the analysis of multispectral images is applied. At performance of recognition taking into account statistical connedtions the joint distributions of a probability density showing the dependence of the general features vector of objects on their classes and on the general vector of conditions at which the set of processed images is received have defining role. Estimations of such distributions are formed by means of a learning sample. The decision on an object class is accepted on a maximum of a posteriori probability counted on the basis of joint distribution. By analogy with the first variant, in the article basic formulas and the scheme, realizing model of complex recognition taking into account mutual statistical connections between patterns of objects are presented. In both variants after performance of operation of recognition a calculation of an estimation of conditional probability of correct recognition is carried out. Resultants indicators of quality of recognition are the full matrix of probabilities and average probabilities of correct recognition and misrecognition.