Semantic segmentation of historical documents via fully-convolutional neural network

Abstract

This paper presents a method for character semantic segmentation in full-text documents from post World War II Czechoslovakia. Unfortunately, standard optical character recognition algorithms have problems to accurately read these documents due to their noisy nature. Therefore we were looking for some ways to improve these unsatisfactory results. Our approach is based on fully-convolutional neural network inspired by U-Net architecture. We are utilizing a synthetic image generator for obtaining a training set for our method. We reached 99.53% recognition accuracy for synthetic data. For real data, we are providing qualitative results.