ON THE FEATURES OF THE FORMATION OF DESCRIPTORS IN THE SIAMESE NEURAL NETWORK

Abstract

The subject of research ̶ the processes of image recognition of handwritten numbers using neural networks. The recognition application is based on the architecture of the Siamese network with neural convolutional subnets. The purpose of the article is to substantiate the choice of N-dimensional vector representations of input images to describe their properties, compare and recognize them. Objective: experimental study of handwritten number image recognition using the architecture of the Siamese neural network. Research methods: direct search method for functions with several variables to determine N-dimensional vector representations of input images. Research results. The definition of N-dimensional vector representations of input images of handwritten numbers is carried out and their characteristics are investigated. An experimental study of image recognition using vector representations of images within the architecture of the Siamese neural network. It is shown that the proposed methods for determining vector N-dimensional representations of input images are robust and have little effect on the number of errors in recognition testing. Images of handwritten numbers from the MNIST test set were used during testing. It is determined that the use of preselected reference representations of the input images can simplify the architecture of the Siamese network. Conclusions. The results obtained in this work can be used in the comparative evaluation and determination of N-dimensional vector representations of different classes of input images in order to recognize them using the architecture of the Siamese neural network.