Receipt date: 
01.02.2019
Year: 
2019
Journal number: 
УДК: 
004.932
Article File: 
Pages: 
82
95
Abstract: 

The article considers one of the approaches to solving the problem of automated detection of defects of the upper structure of the railway track on the basis of image processing methods. The whole problem is broken into a series of sequential subtasks: image preprocessing, detection of rail defects, rail fasteners search, identification of fasteners that are suspected to defect. For preprocessing, an approach based on the use of matrix filters for the convolution operation was used. The solution of the problem of finding rail defects was performed on the basis of the algorithm for finding areas other than a given color within a given rectangular area. Rail fasteners were identified on the basis of the pattern matching algorithm adapted to the characteristics of the subject area. In view of the variety of defects, it was decided to divide the images into two classes: with and without defects, followed by the transmission of images suspicious of the defect for manual processing. The images are divided into four groups:

- mistakes; in such images of bonds or not at all, or they are not completely caught in the perspective due to inaccuracies of the algorithm for finding bonds;

- real defects; on such images of fastening or are absent in view of malfunction, or there is other breakage;

- perfect staples; here the staples images are free of defects and are fully visible;

- other; in such images there are white spots, debris, bonds can be covered with stones, covered with other objects and more.

All non-ideal images are considered suspicious of the defect. A convolutional neural network was used to separate them from the ideal ones. The issues of training such a network are discussed

List of references: 
  1. LeCun, Y. Convolutional Networks for Images, Speech, and Time-Series, in Arbib / LeCun, Y. Bengio ; MIT Press, 1995 – 14с.
  2. Simard, P.Y. Best Practices for Convolutional Neural Networks Applied to Visual Document Analysis / P.Y. Simard, D. Steinkraus, J. Platt // International Conference on Document Analysis and Recognition (ICDAR), IEEE Computer Society. – Los Alamitos. – 2003. – P. 958-962.
  3. Kozin, N.E. Gradual learning the radial neural networks / N.E. Kozin, V.A. Fursov // Computer Optics. – 2004. – № 26. – P. 138-141.
  4. LeCun, Y. Gradient Based Learning Applied to Document Recognition / Y. LeCun, L. Bottou, P. Haffner – IEEE Press, 1998. – P. 46.
  5. Khaykin S. Neyronnyye seti: polnyy kurs. 2-e izdaniye [Neural networks: full course. 2nd edition]: Per. s angl. N.N. Kussul. A.Yu. Shelestova; pod red. N.N. Kussul – M.: Izdatelskiy dom «Viliams». 2006. – 1104 p. (in Russian).
  6. Kolmogortsev M.G. Osipova V.E. Razrabotka modeli opredeleniya mezhpoyezdnykh in-tervalov s ispolzovaniyem apparata iskusstvennykh neyronnykh setey [Development of a model for determining inter-train intervals using artificial neural networks] // Izvestiya Transsiba. Vol.4(28). 2016. – pp. 87-93. (in Russian).
  7. Fisenko V.T., Fisenko T.Yu. Kompyuternaya obrabotka i raspoznavaniye izobrazheniy [Computer processing and image recognition]. – Sankt-Peterburg: Izd-vo SPBGU ITMO. 2008. – 281p. (in Russian).
  8. Reznitskiy M.A. Resheniye zadachi analiza verkhnego stroyeniya puti po ego izobrazhe-niyam [Solving the problem of analysis of the upper structure of the path by its images] // Informatsionnyye tekhnologii i problemy matematicheskogo modelirovaniya slozhnykh sistem. – Vol.16. – Irkutsk: IrGUPS. 2016. – pp. 15-20. (in Russian).
  9. Jahne B. Digital image processing. 6th ed. Springer, 2005. – 639 c.
  10. Hall, E. L. Computer Image Processing and Recognition, Academic Press, New York, 1979. – 811c.
  11. Beloglazov I.N., Tarasenko V.P. Korrelyatsionno-ekstremalnyye sistemy [Correlation-extreme systems]. Sovetskoye radio. 1974. 392 p. (In Russian).
  12. Borisov E.S. O metodakh obucheniya mnogosloynykh neyronnykh setey pryamogo rasprostraneniya [Methods of training multilayered neural networks of direct distribution] – http://mechanoid.kiev.ua/neural-net-backprop2.html (05.11.2018) (in Russian).
  13. Adam: A method for stochastic optimization – https://arxiv.org/abs/1412.6980 (05.11.2018).