Using of a convolutional neural network with changing receptive fields in the tasks of image recognition

Abstract

In the article synthesis procedure of mathematical model features of convolution neural network (CNN) is described. In order to improve the generalization capability of the network the training set is generated by adding a distorted image with changing of CNN receptive fields. This fact differs given procedure from the known procedures. We propose the reduction algorithm of an extended training set and the synthesis algorithm of features for CNN with non-standard receptive fields. The experiments results of the developed algorithms were shown in the article in order to assess of generalization capability changes of the convolution neural network. The experiments were performed with the hardware platform of the “Mechatronics” stand (SPA “Android techniques”, Russia)