M. E. Mazurov – Dr.Sc. (Phys.-Math.), Professor, Russian Economic University n.a. G.V. Plekhanov (Moscow, Russia)

E-mail: mazurov37@mail.ru

The aim of the work is to substantiate the criteria for the accuracy of image recognition in neural networks that are standard ones on McCulloch–Pitts neurons and selective ones using selective neurons. Methods of machine image recognition are listed, it is indicated that the assessment of the recognition accuracy in this case depends on the specific recognition problem and the datasets used for training as a training database in a particular case. The general structure of face recognition is presented. The description of some datasets used in image recognition is given. The deformations of images during real recognition are indicated. Depending on the characteristics of the information components of the system, there are three approaches to the problem of pattern recognition-images: 1) the principle of comparison with the standard; 2) the principle of clustering; 3) the principle of generality of properties.

The structure of feed forward neural networks, standard ones, based on McCulloch-Pitts neurons, and selective ones, based on selective neurons, is presented. When assessing the accuracy of recognition of two-dimensional images in standard and selective neural networks, it is proposed to use a correlation quite convenient criterion.

A description of the types of image recognition in neural networks is given: 1) recognition in a given set of objects – a database; 2) recognition in a given set of objects in the presence of their deformation (for example, in the presence of rotation, shift, resizing of the image); 3) recognition in the presence of noise; 4) recognition with incomplete information about objects in the database. Special databases are used depending on the specificity of recognition caused by a variety of objects. The criterion of accuracy assessment is based on the use of a fundamental property of neurons that are part of a neural network, which is that a neuron is fired when the threshold value of the sum of input signals is exceeded and is not fired if this sum is less than the threshold. This criterion makes it possible to assess the recognition accuracy in feed forward neural networks (standard and selective ones), taking into account the reference dataset used. The proposed criterion for assessing the accuracy allows for a visual analytical and geometric interpretation. The proposed accuracy criterion was used practically to assess the accuracy of image recognition in the form of a set of numbers, to recognize portraits of Russian and Soviet writers and poets, and "smart" recognition of electrocardiograms.

Mazurov M.E. Assessment of accuracy of image recognition in selective neural networks. Neurocomputers. 2020. Vol. 22. No. 5. P. 64–75. DOI: 10.18127/j19998554-202005-06. (in Russian)

1. Forsajt D.A., Pons Zh. Komp'yuternoe zrenie. Sovremennyj podkhod: Per. s angl. M.: ID «Vil'yams». 2004. (in Russian)
2. Zhuravlev Yu.I., Ryazanov V.V., Sen'ko O.V. Raspoznavanie. Matematicheskie metody. Programmnaya sistema. Prakticheskie primeneniya. M.: Fazis. 2005. (in Russian)
3. Zhao W., Chellappa R., Phillips P.J., Rosenfeld A. Face recognition: A literature survey. ACM Computing Surveys. 2003. V. 35. № 4. P. 399–485.
4. Gilani S.Z., Mian A. Learning from millions of 3D scans for large-scale 3D face recognition. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2018. [Elektronnyj resurs] arXiv:1711.05942.
5. Fillips P.Dzh., Flinn P.Dzh., Skraggs T., Bojer K.V., Chang Dzh., Khofman K., Markes Dzh., Min Dzh., Vorek V. Obzor grandioznoj zadachi raspoznavaniya lits. V kn.: Komp'yuternoe zrenie i raspoznavanie obrazov. 2005. S. 947–954.
6. Yin L., Wei X., Sun Y., Wang J., Rosato M.J. A 3D facial expression database for facial behavior research. 7th International Conference on Automatic Face and Gesture Recognition. 2006. P. 211–216.
7. Gupta S., Castleman K.R., Markey M.K., Bovik A.C. Texas 3D face recognition database. IEEE Southwest Symp. on Image Analysis and Interpretation. 2019. P. 97–100.
8. LeCun Y., Bottou L., Bengio Y., Haffner P. Gradient-based learning applied to document recognition. Proceedings of the IEEE. 1998. V. 86. № 11. P. 2278–2324.
9. Khajkin S. Nejronnye seti: polnyj kurs. Izd. 2-e. M.: ID «Vil'yams». 2006. (in Russian)
10. Galushkin A.I. Nejronnye seti. Osnovy teorii. M.: Telekom. 2010. (in Russian)
11. Osovskij S. Nejronnye seti dlya obrabotki informatsii: Per. s pol'skogo I.D. Rudinskogo. M.: Finansy i statistika. 2002. (in Russian)
12. Magnitskij N.A. Nekotorye novye podkhody k postroeniyu i obucheniyu iskusstvennykh nejronnykh setej. Nelinejnaya dinamika i upravlenie. 2001. Vyp. 1. S. 87–98. (in Russian)
13. LeCun Y., Bengio Y., Hinton G. Deep learning. Nature. 2015. V. 521. P. 436–444.
14. Patent № 2597495 RF. Nejron, modeliruyushchij svojstva real'nogo nejrona. M.E. Mazurov. Opubl. 10.09.2016. Byul. № 25. (in Russian)
15. Patent № 2597497 RF. Odnoslojnyj pertseptron na osnove izbiratel'nykh nejronov. M.E. Mazurov. Opubl. 10.09.2016. Byul. № 25. (in Russian)
16. Patent № 2598298 RF. Impul'snyj nejron, blizkij k real'nomu. M.E. Mazurov. Opubl. 20.09.2016. Byul. № 26. (in Russian)
17. Patent № 2597496 RF. Odnoslojnyj pertseptron, modeliruyushchij svojstva real'nogo pertseptrona. M.E. Mazurov. Opubl. 10.09.2016. Byul. № 25. (in Russian)
18. Mazurov M.E. Intelligent recognition of electrocardiograms using selective neuron networks and deep learning. International Conference of Artificial Intelligence, Medical Engineering, Education. Moscow, Russia. 2017. P. 182–198.
19. Mazurov M.E. Nelinejnaya dinamika, pochti-periodicheskoe summirovanie, avtokolebatel'nye protsessy, informa-tsionnoe kodirovanie v izbiratel'nykh impul'snykh nejronnykh setyakh. Izv. RAN. Ser. «Fizicheskaya». 2018. T. 82. № 11. S. 1564–1570. (in Russian)
20. Mazurov M.E. Mekhanizmy invariantnogo pomekhoustojchivogo kodirovaniya v impul'snykh nejronnykh setyakh. Izv. RAN. Ser. «Fizicheskaya». 2020. T. 84. № 1. S. 90–95.