Determination of Expected Output Signals of the Neural Network Model Intended for Image Recognition

摘要

The article is devoted to the problem of further development of the methodology of using neural network models based on a multilayer perceptron for image recognition. It is shown that a significant shortcoming of modern methodologies is the insufficient efficiency of learning neural network models, is associated with the insufficient quality of processing statistical data that are used in the formation of the parameters of the training example. One of the important reasons for this is the lack of correlation between the expected output of the training examples with the closeness of the standards of the classes that should be recognized. It is suggested to improve the quality of the training examples by using the method of neural network coding the magnitude of the expected output signal of the training examples, which will allow taking into account the proximity of the standards of recognized classes in such a signal. A method for determining the encoding of the expected output signal is developed, which provides for estimating the closeness of standards using a probabilistic neural network. The expediency of using a probabilistic neural network is determined on the basis of the low resource intensity of its training. In addition, in the training examples of such a network, the expected output signal can be represented not by a number, but by the name of the expected class. The mathematical apparatus of the method is formed. As a result of numerous experiments it was shown that the application of the developed method makes it possible to reduce the number of training iterations in 1.3-1.5 times to achieve an allowable error of training within 1%. It is shown that the prospects for further research are to adapt the method to neural network systems used in industries in which the computational complexity of learning is of critical importance.