APPLICATION FEATURES OF ARTIFICIAL NEURAL NETWORK FOR THE RAILWAY CAR NUMBER PLATE RECOGNITION

Abstract

The article analyzes the systems of automatic identification of rolling stock during movement. In recent decades, both in Ukraine and in CIS countries, much attention has been paid to the development of systems for the automatic reading of information from moving units using special sensors that attach to the body of the railway cars. The most common of these sensors are passive radio tags or RFID tags. The implementation of such systems on the railways of several countries allows to solve many problems of automation of processes of control and management on railway transport. In the article the advantages and disadvantages of such systems are highlighted, the optical method of recognition of inventory numbers of railway cars is given attention. Such systems for Ukrainian railways have been insufficiently vandal-protected, as indicated by the collapse of the SAI RS project, after systematic damage to the sensors located on the railway cars. Recently, there has been a tendency in Ukraine and abroad for the use of the optical method and machine vision systems for tracking the movement and identification of moving railway units. The advantage of this method is that there is no need for rolling stock sensors and the ability to constantly improve the machine vision system software. At the same time, the main difficulty of using this method in railway operations is damage to the railway car number. The peculiarities of the use of artificial neural networks for optical method of the number plate recognition on the bodies of railway cars are investigated, the structure of neural network is grounded and the neural network is built for the problem of recognizing the railway car number, and the algorithm of learning of the neural network in the Matlab environment is developed. A direct error propagation method was used to directly train the neural network to minimize the error of the multilayer neuron. The idea behind this method is to propagate the error signals from the network outputs to its inputs in the direction opposite to the direct propagation of the signals in normal mode. The neural network recognized the number of the test image. Thus, even on a highly simplified neural network with three hidden layers, we were able to programmatically achieve the generalization effect.