Character Recognition for Tibetan Ancient Books Based on Deep Learning Approaches

In this paper, a novel method for Tibetan ancient book image character recognition based on deep neural networks (CNN and RNN) was proposed. To solve the problem of ultra-long text line ancient book recognition, the ultra-long text is dynamically split into multiple sub-strings with the sliding window based Tibetan ancient book text line splitting technology, and the adjacent sub-string overlapping characters are processed based on character location information. The string recognition model of Tibetan ancient books based on ResNet[37] (Residual Network) and Bi-LSTM[38] (Bidirectional Long Short-Term Memory) recurrent neural network (RNN) was constructed to solve the problem of character recognition of ancient books with poor image quality, serious adhesion of adjacent characters, and large overlap between up and down lines. Moreover, Through experimental comparison and analysis, the macro average accuracy of our proposed method in modern Tibetan publication print recognition has reached 97.27%, which is 12.01 and 10.95 higher than TH-OCR (85.26%) and UTibet-OCR-2008 (86.32%) respectively. The average accuracy rate of the proposed method for the recognition of Tibetan ancient woodcut characters has reached 92.60%, and the average accuracy rate for the recognition of Tibetan ancient handwritten characters has reached 85.47%, which demonstrates that our proposed method has achieved an important breakthrough in the field of Tibetan ancient character recognition at home and abroad, and has been open to the public for free trial.