I.K. Belova - Ph. D. (Phys.-Math.), Kaluga branch of the Bauman MSTU E.O. Deryugina - Ph. D. (Eng.), Kaluga branch of the Bauman MSTU A.V. Ermolenko - Post-graduate Student, Kaluga branch of the Bauman MSTU. E-mail: syvorova_eo@mail.ru

One of the common problems of vision systems is the identification and classification of objects in the scene image. The problem of the search of the desired sample is relevant for a wide range of applications for the establishment of the authors handwriting on it bitmap, and for the determination of the desired object on the primitive image of the cosmos, including from a stereo pair of video cameras in the synthesis of three-dimensional image. The functioning of the neural network model is based on the critically informative feature space, which allows to solve the problem of separability of sets by a hyperplane. The task of building the feature space is the definition of the basic elements of reform, in order to reduce the invariance of the objects covered in signs. Typically, images of objects in the real scene as a two-dimensional (eg, text) as well as three-dimensional, are noisy character, forcing to access algorithms weakly sensitive to interference. The structure of the machine classifier is a synthesis of algorithms of construction signs and solver. Studied aspects of contour analysis methods for solving the problem of identification of segmented images. The authors have developed an original method of finding a segment of the sample on the stage. The comparative analysis of the discrete Fourier transform and a two-dimensional Radon transform. The architecture of the neural network identification of the device, a modified method of contour analysis and image pre-processing filters. The above article aspects of search engine algorithms and restore arbitrary ellipse characteristics allow us to conclude that the de-termination of the guaranteed characteristics of the ellipse in the frame is determined by a five-dimensional Radon algorithm. However, the practical application of this algorithm is not efficient due to the high algorithmic complexity. frame analysis based on discrete Fourier transform provides a very fast and high quality results, but only for a class of problems in which the parameters are accurately known desired pattern. Also, the result may vary depending on the parameters of pre-segmentation of the original image. Direct neural network algorithm algorithmically complex and loyal to the input conditions. But the real load of image noise and generally similar to the desired object does not achieve significant reliability of the result. The proposed modified Radon algorithm with neural network filtering can significantly reduce the algorithmic complexity of the classic five-dimensional algorithm. This preserves the accuracy of the identification is not lower than a discrete Fourier transform. However, unlike a Fourier modified Radon algorithm requires no prior frame segmentation and clear characteristics of the desired sample. Therefore, it can be applied within a broader framework of input conditions that would be necessary to solve the problems of identification of graphic images to view images of the scene.

 

1. Vinogradov A.N., Kalugin F.V., Nedev M.D., Pogodin S.V., Talalaev A.A., Tishhenko I.P., Fralenko V.P., KHachumov V.M. Vydelenie i raspoznavanie lokalnykh obektov na aehrokosmicheskikh snimkakh // Aviakosmicheskoe priborostroenie. 2007. № 9. S. 39−45.
2. Belova I.K., Derjugina E.O. Rekurrentnaja nejjronnaja set s upravljaemymi sinapsami dlja raspoznavanija rukopisnogo svjaznogo teksta // EHlektromagnitnye volny i ehlektronnye sistemy. 2013. T. 18. № 10. S. 36−39.
3. Ermolenko A.V., Kozlichenkov A.A., Korljakova M.O. Vybor vkhodnogo nabora priznakov dlja nejjrosetevojj identifikacii pocherka // Nauchnaja sessija MIFI 2007. IX Sb. nauchnykh trudov Vseros. nauchno-tekhnich. konf. «Nejjroinformatika-2007»: V 3-kh chastjakh. CH. 1. M.: MIFI. 2007. S. 42−50.
4. Ermolenko A.V., Korljakova M.O., Filimonkov A.A. Formirovanie nejjrosetevojj modeli geoinformacionnogo datchika // Izvestija Instituta inzhenernojj fiziki. 2011. № 1(19). S. 25−30.
5. Jun Xie, Pheng-Ann Heng, Mubarak Shah Shape matching and modeling using skeletal context // Pattern Recognition. May 2008. V. 41. № 5. P. 1429−1862. P. 1756−1767.
6. Kazanovich JA.B. Segmentacija izobrazhenijj s pomoshhju dinamicheskikh nejjronnykh setejj // Nauchnaja sessija MIFI-2008. KH Vseros. nauchno-tekhnich. konf. «Nejjroinformatika-2008»: Lekcii po nejjroinformatike. CHast 2. M.: MIFI. 2008. S. 37−97.
7. Zaprjagaev S.A., Sorokin A.I. Programmnaja obolochka dlja poiska primitivov na izobrazhenii // Vestnik VGU. Ser. Sistemnyjj analiz i informacionnye tekhnologii. № 2. 2008. S. 37−47.
8. ZHuravlev JU.I. Ob algebraicheskom podkhode k resheniju zadach raspoznavanija ili klassifikacii // Problemy kibernetiki. M.: Nauka. 2005. № 33. S. 5−25.
9. Makarov I.M., Lokhin V.M., Manko S.V., Romanov M.P. Iskusstvennyjj intellekt i intellektualnye sistemyupravlenija. M.: Nauka. 2006. 333 s.