Structural-parametric synthesis of convolutional neural networks under the presence of noise in the input data

Abstract

The purpose of this work is to create and develop the approach on structural parametric synthesis of convolutional neural network to receive the unique CNN architecture with the good image recognition accuracy. The paper deals with the methods of processing and classification of graphic images using convolutional neural networks and mathematical algorithms for their support. Using researches, there shown that for the proper use of such system it’s requires compliance with special technical conditions. Today, in modern convolutional neural networks for the independent processing of graphic data there is a problem of lack of accuracy in the selection of special criteria. The urgency of this problem over time is only increasing due to the proliferation of the problem of digital identification. In order to increase the accuracy of the results of the work, there designed system includes the algorithm of input data preparation, generating the neural network architecture and configuration its global and local parameters with means of structural parametric synthesis algorithms. Also, there were done relative surveys and tests as well as implemented all the algorithms by means of programming.