Konvoluční neuronové sítě pro klasifikaci písma

Abstract

This diploma thesis explores possibilities of using convolutional neural networks for visual recognition of asian languages (chinese, japanese and korean) by the appearance of it’s characters. The main goal of this work is to create a model for localization and classification of these languages’ text in in natural scene images. The work contains design and implementation of a synthetic data generator for improving the resulting model. There are also experiments with different architectures, learning methods and hyperparameters configurations with the goal to find an optimal solution.