|Article title||INVESTIGATION OF CONVOLUTION NEURAL NETWORK WHICH WAS TRAINED BY THE METHOD OF USE OF NON-STANDARD RECEPTIVE FIELDS IN THE PROCESS OF IMAGE RECOGNITION|
|Section||SECTION II. COMPUTER ENGINEERING AND COMPUTER SCIENCE|
|Month, Year||07, 2015 @en|
|Abstract||In this article we describe the method of synthesis of mathematical model parameters of convolution neural network (CNN), which differs from the known fact that for to improve the generalization capability of the network, training set is generated by adding a distorted image by changing the receptive fields of the CNN. As a result, the same pattern is seen with the network differently. Considerable attention is paid to the analysis of the results of experiments on the recognition of objects in vision systems "Mechatronics" stand and the investigation of robustness of convolution neural network to the noise. An adaptation of backpropagation algorithm to CNN with receptive fields of non-standard form is details described. It is shown that the use of non-standard forms of receptive fields allows you to distort the input patterns and thereby expand the training set, increasing the generalizing of the network. The practical testing of the proposed method of parameters synthesis of the CNN model was conducted on technical platform of "Mechatronics" stand (manufacturer SPA "Android technics", Russia). Convolution neural network trained using the extended sampling for recognition 10 object classes was integrated in the vision system of this stand. The structure of convolutional neural network was described in this article. This network has two types of layers: convolutional and subsabmping layers. Given the importance of the robustness of the vision system to noises, experiments were conducted with white noise. The results of an experimental study of CNN which was trained proposed method and without it were described in this article. The boundaries of the permissible noise level which does not affect the quality of recognition was set.|
|Keywords||Convolutional neural networks; pattern recognition; receptive fields; training set; white noise.|
|References||1. Gorban' A.N., Rossiev D.A. Neyronnye seti na personal'nom komp'yutere [Neural network on a personal computer]. Novosibirsk: Nuka, 1996, 278 p.
2. Dileep G., Hawkins J. A Hierarchical Bayesian model of invariant pattern recognition in the visual cortex, In Proceedings of the International Joint Conference on Neural Networks, 2005, Vol. 3, pp. 1812-1817.
3. LeCun Y., Bottou L., Bengio Y., Haffner P. Gradient-based Learning Applied to Document Recognition, Proceedings of the IEEE, 1998, Vol. 86, pp. 2278-2324.
4. Khaykin S. Neyronnye seti. Polnyy kurs [Neural network. Full course]. 2nd ed: Translation from English, 2008 g. Moscow: Izd-vo Vil'yams, 2008. 1104 p.
5. Bishop C.M. Pattern Recognition and Machine Learning. New York: Spring Street, 233, LLC, USA, 2009, 738 p.
6. Hassoun M. Fundamentals of Artificial Neural Networks. A Bradford Book, 2003, 511 p.
7. Nemkov R.M., Mezentseva O.S. Dinamicheskoe izmenenie vosprinimayushchikh svoystv svertochnykh neyronnykh setey kak obuchenie s shumom i ego vliyanie na obobshchayushchuyu sposob-nost' [Dynamic change of the perceiving properties of convolutional neural networks as training with noise and its impact on the generalization performance], Neyrokomp'yutery: razrabotka, primenenie [Neurocomputers: development, application], 2015, No. 2 (63). pp. 12-18.
8. Nemkov R., Mezentseva O. The Use of Convolutional Neural Networks with Non-specific Receptive Fields, The 4th International Scientific Conference: Applied Natural Sciences. Novỳ Smokovec, 2013, pp. 284-289.
9. Nemkov R.M. Razrabotka chislennogo algoritma reduktsii rasshirennogo obuchayushchego mnozhestva, generiruemogo izmeneniem retseptivnykh poley svertochnoy neyronnoy seti [Development of numerical algorithm of reduction of the expanded training set generated by a change receptive fields of convolutional neural networks], Materialy II Vserossiyskoy nauchno-tekhnicheskoy konferentsii «Studencheskaya nauka dlya razvitiya informatsionnogo obshchestva» [Materials of II Russian scientific-technical conference "Student science for development of the information society"]. Stavropol', 2015, pp. 262-264.
10. Nemkov R.M. Metod sinteza parametrov matematicheskoy modeli svertochnoy neyronnoy seti s rasshirennym obuchayushchim mnozhestvom [Method of synthesis of parameters of the mathematical model of convolutional neural network with advanced learning lots], Sovremennye problemy
nauki i obrazovaniya [Modern problems of science and education], 2015, No. 1.
11. Nemkov R.M., Mezentseva O.S. Dinamicheskoe izmenenie vosprinimayushchikh svoystv svertochnykh neyronnykh setey kak obuchenie s shumom [Dynamic change of the perceiving properties of convolutional neural networks as training with noise], Parallel'naya komp'yuternaya algebra i ee prilozheniya v novykh infokommunikatsionnykh sistemakh:
sbornik nauchnykh trudov I Mezhdunarodnaya konferetsii [Parallel computer algebra and its applications in new information and communication systems: proceedings of the I International conference]. Stavropol', 2014, pp. 539-548.
12. Nemkov R. Dynamical Change of the Perceiving Properties of Neural Networks as Training with Noise and Its Impact on Pattern Recognition, Young Scientists’ International Workshop on Trends in Information Processing (YSIP) – 2014. Available at: http://ceur-ws.org/Vol-1145/paper4.pdf/ (accessed 26 July 2015).
13. Katalog produktsii NPO «Androidnaya tekhnika» [Product catalog NGOs "Android technology"]. Available at: http://npo-at.com/products/katalog-produktsii/ (data obrashcheniya: accessed 26 July 2015).
14. Nemkov R.M., Mezentseva O.S. Razrabotka modulya obucheniya svertochnykh neyronnykh setey dlya kompleksa raspoznavaniya izobrazheniy s kamery mobil'nogo robota [Development of module training convolutional neural networks for image recognition with a camera of a mobile robot], Fundamental'nye i prikladnye aspekty komp'yuternykh tekhnologiy i informatsionnoy bezopasnosti: materialy trudov I Vserossiyskoy nauchno-tekhnicheskoy konferentsii [Fundamental and applied aspects of computer technologies and information security: materials of I all-Russian scientific-technical conference]. Rostov-na-Donu, 2015, pp. 463-465.
15. Lagunov N.A., Mezentseva O.S. Vliyanie predobrabotki izobrazheniy na kachestvo obucheniya neyronnoy seti dlya ikh raspoznavaniya [The influence of pretreatment images at the quality of training the neural network for recognition], Vestnik Severo-Kavkazskogo gosudarstvennogo
tekhnicheskogo universiteta [Bulletin of North Caucasus state technical University], 2015, No. 1 (46), pp. 21-25.
16. Gonsales R., Vuds R. Tsifrovaya obrabotka izobrazheniy [Digital image processing]. 3rd ed. Moscow: Tekhnosfera, 2012, 1104 p. ISBN 978-5-94836-331-8.
17. Sermanet P., LeCun Y. Traffic sign recognition with multi-scale convolutional networks, In Proceedings of the International Joint Conference on Neural Networks (IJCNN), August, 2011.
18. Belyy shum [White noise]. Available at: https://ru.wikipedia.org/wiki/Belyy_shum (accessed 26 July 2015).
19. LeCun, Y., Huang F.-J., Bottou L. Learning methods for generic objects recognition with invariance to pose and lighting, in Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR’04), Vol. 2, pp. 97-104. Los Alamitos, CA, USA: IEEE Computer Society, 2004.
20. Ciresan D., Meier U., Schmidhuber J. Multicolumn Deep Neural Networks for Image Classification, In Proceedings of the 2012 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), CVPR '12. pp. 3642-3649, Washington, DC, USA, 2012. IEEE Computer Society.