IMPROVING THE EFFICIENCY OF THE CONVOLUTIONAL NEURAL NETWORK USING THE METHOD OF INCREASING THE RECEPTIVE FIELD

Abstract

Listen

Translate article

In many pattern recognition problems solved using convolutional neural networks (CNN), one of the important characteristics of network architecture is the size of the convolution kernel, since it coincides with the size of the maximum element that can act as a recognition sign. However, increasing the size of the convolution kernel greatly increases the number of tunable network parameters. The method of effective receptive field was first applied on AlexNet in 2012. The practical application of the method of increasing the effective receptive field without increasing convolution kernel size is discussed in this article. A presented example of a small network designed to recognize a fire in apicture demonstrates the use of an effective receptive field which consists of a stack of smaller convolutions. Comparison of a original network with a large convolution core and a modified network with a stack of smaller cores shows that, with equal network characteristics, such as prediction accuracy, prediction time, the number of parameters in the network with an effective receptive field, the number of tunable parameters is significantly reduced.