Recognition of handwritten letters and numbers using deep learning neural networks (In Ukrainian)

This article discusses several variants of convolutional neural network architecture for recognizing isolated handwritten Latin or Ukrainian letters and numbers that have been trained using synthetic datasets of two types, built on the basis of a set of handwritten and italic fonts or a CoMNIST dataset. A comparison of the recognition results of several variants of images containing handwritten letters and numbers using models with different architectures showed that an increase in the number of convolutional layers leads to a decrease in the frequency of erroneous character recognition. The size of the training dataset significantly affects the reliability of character recognition. The data sets used in the paper contained from 192 to 2304 samples per class. The limit on the number of samples per class that provided acceptable recognition accuracy was about 1,500 images per class. Reducing the sample by reducing the number of samples per class leads to a significant decrease in recognition accuracy (from 90% of the accuracy of recognizing elements of real labels to 40-60% with a 4-fold decrease in the sample size).