Recognition of Chinese Text in Historical Documents with Page-Level Annotations

In historical documents, Chinese characters can be categorized into more than 8000 categories and are hard to recognize by directly applying classic methods. Furthermore, the lack of well-labelled data makes it difficult to recognize them using deep learning based methods. In this paper, we introduce a historical Chinese text recognizer that is trained by data labelled in page-level without the alignment of each text line. We propose Adaptive Gradient Gate AGG) to reduce the influence of misalignments between text line images and labels. With the help of the AGG, the error rate of the proposed text recognizer can be reduced by over 35%. Furthermore, we find that the implicit language model induced by a text recognition network with Convolutional Neural Networks(CNNs) and Connectionist Temporal Classification (CTC) as one important factor to achieve high recognition performance. Lastly, we compare the proposed recognizer with the three leading optical character recognition systems on the market. Our comparison reveals that the proposed recognizer outperforms other character recognition systems significantly.