USING MACHINE LEARNING MODELS AND DEEP LEARNING NETWORKS FOR HANDWRITTEN NUMBERS AND LETTERS RECOGNITION

Abstract

This article delves into the utilization of a multitude of classification algorithms and deep learning neural networks for the purpose of identifying handwritten letters in photographs or during manual input. A total of eight variations of recognition technology were subjected to analysis and testing, the incorporation of classifiers sourced from the Scikit-learn package, as well as the utilization of deep learning neural networks, are both integral components of the study at hand. In order to construct and train these neural networks or train classifiers, we opted for well-established and comprehensive databases, namely the base of handwritten digits MNIST, the base of handwritten letters of the Latin alphabet EMNIST, and the CoEMNIST dataset for Cyrillic characters. Two types of neural networks were taken into consideration, namely sequential and convolutional. The neural networks were trained through the utilization of varying numbers of epochs. The recognizable images were resized to 28x28 (784 cells can be represented in one dimension). The preprocessing of images (such as filtering and scaling) was undertaken utilizing the OpenCV library. Upon the construction of recognition models through the application of Scikit-learn classifiers or neural networks, the accuracy of recognizing all handwritten digits through the test program ranged between 98% and 100% (with one or no errors in the identification of 50 handwritten digits).The accuracy of recognition of handwritten letters of the Latin alphabet using the MLP classifier turned out to be somewhat worse - at the level of 93-96%. The accuracy of recognition of handwritten Cyrillic letters using the SVC classifier or the CNN network was lower - at the level of 90-92%.