融合字符结构特征的多任务老挝语文字识别研究

A Multi-task Approach to Lao Character Recognition with Structural Features

老挝语属于资源稀缺型语言,直接从互联网中获取老挝语文本语料较为困难,老挝语文字识别研究可在有限的图片文本资源中获取更多的老挝语文本语料。在开展老挝文字光学字符识别的研究工作中,针对老挝单字符误切分、上/下位元音以及音调识别位置存在偏差和相似老挝字符的识别问题,该文研究了老挝字符书写等级和下位辅音,提出一种有效融合老挝字符结构特征的多任务字符识别方法。首先,利用深度残差网络提取字符图片中的老挝字符结构特征,通过边框回归矫正单字符包围框;其次,将已矫正切分结果和提取的字符特征作为联合输入,通过双向长短时记忆网络预测老挝字符序列,利用连接主义时间分类对预测结果进行序列对齐;最后,根据老挝字符固定组合优化模型预测结果。实验结果表明:该方法可以精确识别已切分的老挝字符序列,字符错误率指标低至13.06%。

Focused on the Optical Character Recognition of Lao script,this paper investigates the problems of Lao characters mis-segmentation,the misperception of hypernym/hyponym vowels and tone,and the confusion of simi-lar Lao characters.According to the writing scheme and the hypo consonant of Lao characters,this paper proposes a multitasking character recognition to effectively integrate the structural features of Lao characters.The model ex-tracts the structural features of Lao characters from character pictures via Deep Residual Network,and corrects the single character bounding box through Bounding Box Regression.Then,the corrected segmentation results and ex-tracted character features are input jointly into Bi-directional Long-Short Term Memory network to identify the Lao character sequence,and the sequence alignment is completed by the Connectionist Temporal Classification.Finally,the result is predicted by the fixed combinatorial optimization model of Lao characters.The experimental result shows the method can reduce the Character Error Rate to 13.06%.