Augmentation of Handwritten Slovak Diacritical Letters Using GAN Networks

This paper addresses the challenge of the limited availability of quality datasets for handwritten Slovak letters with diacritical marks, which are crucial OCR applications. Manually expanding these datasets is time-consuming and resource-intensive, so augmenting them using Generative Adversarial Networks (GANs) was explored. DCGAN, cDCGAN, and AC-DCGAN architectures to generate synthetic images of Slovak letters with diacritical marks and incorporated these into the existing EMNIST dataset were utilized. The experiments showed that adding these generated images significantly improved CNN accuracy in classifying Slovak letters, with AC-DCGAN providing the best results. These findings indicate that GAN-generated synthetic data is an effective way to address data scarcity and enhance machine learning model performance, particularly for languages with unique character sets like Slovak.