A Feature Refinement Patch Embedding-Based Recognition Method for Printed Tibetan Cursive Script

Recognition of Tibetan cursive scripts has important applications in the field of automated Tibetan office software and ancient document conservation. However, there are few studies on recognition of Tibetan cursive scripts. This paper proposes a printed Tibetan cursive script recognition method based on feature refinement patch embedding. Firstly, the feature refinement patch embedding module (FRPE) serializes the line text image of feature sequences. Secondly, a global modeling of feature vectors is carried out by using a single transformer encoder. Finally, the output of the recognition result is decoded by using a fully connected layer. Experimental results show that, compared with the baseline model, the proposed method improves the accuracy by 9.52% on the dataset CSTPD, a database containing six Tibetan cursive fonts. Moreover, it achieves an average accuracy rate of 92.5% on the dataset CSTPD. Similarly, it also works better than the baseline model on Tibetan text recognition synthetic data for natural scene images.