乌梅印刷多字体藏文文本的检测与识别

Study on Detection and Recognition of Tibetan Wumei Printing Multi-Fonts Text

随着藏文信息处理技术的发展,藏文乌金字体的识别取得了很好的效果,但藏文乌梅字体由于书写风格差异大,检测和识别难,目前的乌梅字体识别仅限于以字丁识别、单一字体为主。近几年随着计算机字体的丰富,出现了乌梅印刷多字体文本。为了准确识别这类文本,文章基于中英文的预训练模型DBNet开展藏文文本检测,以ResNet-50为骨干网络的CRNN和SRN两种不同编码-解码方式开展端到端的乌梅印刷多字体文本识别,并以实验测试两种模型的识别结果。实验表明,当训练和测试所用字体一致时两个模型的识别效果相当;使用不在训练集中的另外8种乌梅字体进行测试时,SRN识别算法相比CRNN在TCR、TDR和LRA三个评价指标上分别提升0.5363%、1.7681%和3.4875%,表现出更强的泛化能力。

With the development of Tibetan information processing technology,the recognition of the Tibetan Wujin font has achieved good results.However,due to the significant difference in writing style and difficulties in the detection and recognition of the Tibetan Wumei font,the current Wumei font recognition is only capable of recognizing the character and single font.In recent years,with the enrichment of computer fonts,there has been a multi-font text printed by Wumei.To recognize such texts,in this paper,we carried out a Wumei text detection based on DBNet,a pre-trained model in Chinese and English,and an end-to-end multi-character Wumei text recognition using CRNN and SRN,two different encoding and decoding methods,with ResNet-50 as the back-bone network.And the recognition results of the two models are examined.The experiment results show that when the fonts used in training and testing are consistent,the recognition effect of the two models is comparable.While,when examined using other eight Wumei fonts which were not included in the trainning set,compared with CRNN,the SRN recognition algorithm improved by 0.5363%,1.7681%,and 3.4875%on the three evalua-tion indexes of TCR,TDR,and LRA,respectively,showing a better generalization ability.