An Effect of Binarization on Handwritten Digits Recognition by Hierarchical Neural Networks

Abstract

The influence of binarization on the recognition of handwritten digits has been investigated. The main methods of converting images from grayscale to black and white images are analyzed. The binarization process is applied to images of handwritten digits. Both the Convolutional neural network (CNN) and the multilayer perceptron (MLP) architectures are proposed for an automated recognizing the handwritten digits. The CNN and MLP are trained on the raw data of images. Essential characteristics are obtained by the method of feature extraction from binarized images and then are used for the CNN and MLP training. The influence of the binarization threshold on network recognition errors is analyzed. A comparison of the recognition accuracy by the neural networks for different algorithms is carried out. It was estimated that the minimum error of combining the proposed networks does not exceed 99.86%. A new approach to the design of a hierarchical network composed of a network trained on basic images of digits and images of digits obtained by binarization is proposed. The hierarchical neural networks with an accuracy of 99.61% on the MNIST dataset are proposed.