Convolutional Networks for Historic Text Recognition

The aim of this work is to create a tool for automatic transcription of historical documents. The work is mainly focused on the recognition of texts from the period of early modern times written primarily using font called Fraktur. The problem is solved using convolutional neural networks with addition of Spatial Transformer Network. The solution also includes implemented generator of artificial historical texts. The proposed neural network was trained on a dataset created by this generator and for evaluation real historical texts were used. On the real historical dataset, the network achieved 81.8 % of correctly recognized characters. The benefit of this work is the finding that it is possible to train the neural network on artificial data to be able to recognize real historical texts.