Information technology for recognition of handwritten ukrainian letters and numbers using synthetic data sets

Abstract

The paper considers several variants of the architecture of convolutional neural networks for recognizing isolated handwritten digits and Ukrainian letters. The results of recognizing different images containing letters and numbers were compared on models with different architectures. Several variants of rather complex architectures of neural networks were considered. Research was conducted with VGG16 and VGG19 architectures, ResNet or ResNetV2, MobileNet or MobileNetV2, DenseNet. The possibility of learning convolutional neural networks with the help of a synthetic data set built on the basis of handwritten or cursive fonts is shown. The size of the training data set significantly affects the reliability of character recognition. The data sets used in the work correspond in volume to the well-known EMNIST Letters dataset. The lower limit of the sample size, which provides acceptable recognition accuracy, was about 1500 characters per class. Reducing the sample by reducing the number of symbols per class leads to a significant decrease in recognition accuracy (from 90-100% accuracy of recognizing elements of real inscriptions to 40-60% with a 4-fold reduction in the sample size). Thus, when using deep neural networks to recognize letters or numbers, the reliability of recognizing elements of real inscriptions depended primarily on the size of the training data set. The accuracy of recognition of the test data set after training all variants of the models was quite high - 97-98% and higher. However, the generation of training data sets of small size — 300-500 images per class - practically did not provide reliable recognition. In general, when comparing the achieved recognition accuracy of real images and the model training speed, the best performance was provided by the DenseNet or ResNetV2 family models. Experiments with changing the optimization algorithm compared to Adam did not give any improvement in the accuracy and reliability of recognizing real samples. Increasing the number of model training epochs beyond the specified one also did not change the results.