An End-to-End Model for Printed Uyghur Text Recognition

Abstract

The recognition method based on the end-to-end model has a wide range of applications in Chinese and English text recognition, but there are few researches of Uygur text recognition. Using the end-to-end method can effectively avoid the wrong recognition problem caused by the mis-segmentation of Uyghur letters. Based on the Transformer model, we propose a model named EfficientNet-Transformer for printed Uyghur text recognition. By replacing the SE attention of EfficientNet with Triplet attention, the computing ability of the network for spatial and channel attention is improved. The encoder of the original Transformer model is replaced by an improved EfficientNet, which makes the model simpler and fewer parameters. The dataset is expanded by synthesizing words through Uyghur syllable rules, and comparative experiments are carried out on this dataset by using our model with other’s RNN-based model. Experiments show this model is superior to others models in Character error rate, recognition speed and space occupation.