Preview

PDF, English - main document Download (3MB) | Lizenz: Building and Improving an OCR Classifier for Republican Chinese Newspaper Text by Henke, Konstantin underlies the terms of Creative Commons Attribution-NonCommercial-ShareAlike 4.0

Abstract

This work presents methods and results of an initial step towards full text extraction from a Republican Chinese newspaper. My basis is a small fraction of the image corpus for which text ground truth exists. I introduce a character segmentation method which produces over 90,000 labeled images of single characters. Then I pre-train a GoogLeNet classifier as an OCR model on character images extracted from font files and randomly augmented on the fly, whereafter I fine-tune it on the previously segmented character images. I show that the pre-training step is able to increase OCR accuracy from 95.49% to 96.95% on the test set and finally, how post-processing using a masked language model corrects up to 16% of remaining errors, increasing accuracy on the test set to 97.44%.