乌金印刷多字体藏文的文本检测与识别

Wujin printing multi-fonts text detection and recognition in Tibetan

为解决乌金印刷多字体藏文的文本识别以字丁识别为主、识别字体类别单一或较少、无法实现端到端的藏文文本行识别等问题,根据藏文文字的书写特点,在基于分割的文本检测方法DBNet上,对比在MobileNetV3和ResNet34两种骨干网络下CRNN、Rosetta和RARE这3种端到端的文本识别算法;提出一种将常用74个藏文字符作为端到端文字识别的转录字典策略,提出一个针对藏文文本识别的评价指标。实验结果表明,以ResNet34为骨干网络的CRNN文本识别方法在测试集上的综合表现最佳。

To solve the text recognition of multi-font Tibetan texts in Wujin is mainly based on Tibetan-Ding and single or few font types,and it is unable to achieve the end-to-end Tibetan text line recognition,according to the writing characteristics of Tibetan text and a segmentation-based text detection method DBNet,three end-to-end text recognition algorithms CRNN,Rosetta and RARE under the two backbone networks of MobileNetV3 and ResNet34 were compared.A transcription dictionary strategy using commonly 74 Tibetan characters as end-to-end text recognition and an evaluation index for Tibetan text recognition were proposed.Experimental results show that the CRNN text recognition method with ResNet34 as the backbone network has the best comprehensive performance on the test set.