Deep learning based tangut character recognition
Guangwei Zhang, Xiaomang Han · 2017
The Tangut script, a logographic writing system, was used for writing the extinct Tangut language of the West Xia Dynasty. The huge amount of Tangut historical documents are being mainly recognized by Tangut experts manually, because the Tangut language has not been used since 16th century and it was impossible to recognized automatically in the past. With the help of deep learning, we build an end-to-end Tangut character recognition system to reduce the labor of Tangut experts. The high accuracy of a deep learning system for character recognition is essentially guaranteed by a large training dataset of well-labeled data. We construct a training dataset containing more than 100,000 labeled Tangut images, which is used for training a deep convolutional neural network (DCNN) to recognize Tangut characters. The Tangut images in the training dataset are from Tangut historical documents and they are labeled in a cluster-and-label way to reduce the human efforts. Based on the training dataset, the validation accuracy of the DCNN is more than 94% according to our experiments. We will release the training dataset for further study and construct an OCR system for transcribing Tangut historical documents automatically in the future.