Research on Tibetan character recognition based on model CRNN

Abstract

In recent years, CRNN has been widely used in computer vision and has achieved remarkable results in the direction of text recognition. CRNN is a convolutional recurrent neural network structure, which is mainly used in image sequence recognition problems. The CRNN network model implements variable-length verification, combining CNN and RNN networks, using a bidirectional LSTM cyclic network for time series training, and then introducing a CTC loss function to recognize variable-length sequence texts. In the field of Tibetan text recognition, based on end-to-end recognition, it is usually to recognize a line of text. Due to the special structure of Tibetan syllables, the components of Tibetan characters can be split, and end-to-end recognition can be applied to the study of Tibetan single-character recognition. In this paper, a new split-based method is used for end-to-end recognition of single characters in Tibetan ancient books and single characters in Tibetan handwriting using the CRNN model, and a good recognition effect is achieved. It provides a new method for Tibetan character recognition research.