A. F. Ulasen, G. A. Kocur

In article the technique of an estimation of a system effectiveness raspo-znavanija objects under their optical images is considered. In a technique consecutive carrying out of researches for three types нейросетей is provided: multilayered персептрона, сверточной a neural network and сверточной a neural network with multilayered нейросетевым the filter. For each type нейросети the efficiency estimation is spent at various diskret-nosti images ВЦ in training sample, in the conditions of change of conditions of metrological visibility and loss of a part of the image of air object. As the decision of system of recognition in this case depends from mno-gih random factors, an indicator characterising efficiency, you-brana probability of correct and erroneous decisions on purpose type. And considering that the recognition system is intended for use in ЗРС and ЗРК to the small range, the second indicator chooses recognition time. The technique essence consists that training and selection of internal pases-rametrov нейросетевых qualifiers is made on training sample in which images of the purposes with certain step-type behaviour from-menenija a foreshortening are presented. Then the probability of correct recognition and a network operating time on sample in which images of the purposes in all range of change of a foreshortening are presented is defined. By results of the spent estimation, for all types нейросетей according to criterion gets out most ratsio-nalnaja step-type behaviour of change of a foreshortening of images of air targets in obu-chajushchej to sample. Results of researches spent on model on developed meto-dike have allowed to prove possibility of automatic recognition from-brazheny three classes of air targets under the information from an optiko-electronic source on distance to 20 km, with probability is not worse 0,91-0,88 during no more than 6 seconds. Scientific results of work can be used at testing again created нейросетевых recognition systems.