格式化约束的时间戳文字识别网络

Time Stamp Recognition Network Constrainted by Formated Information

文字识别是深度学习网络的重要应用领域,主流算法基于光学信息预测自然场景文字。然而在一些特定领域的文本对象上,额外的关键特征将会进一步提高文字识别算法的准确性。在安防监控领域,画面中的时间戳文本拥有格式规范、限定数值范围等特点,根据这一特点,对时间戳文本识别网络进行了研究,提出一种时间戳信息约束机制,融合文本语义约束信息和光学特征达到识别规范文本的效果,增强输出时间戳文本的格式规范性和数值合理性。在全匹配率、编辑距离等标准上全面超过基于光学特征的经典文字识别算法。

Text recognition is an important application field of in-depth learning network.Mainly used algorithms predict text in natural scenes based on optical information.However,on some domain-specific text objects,additional key features will further improve the accuracy of text recognition algorithms.In the field of security monitoring,timestamp text in pictures has the characteristics of standard format and limited value range.This paper studies the timestamp text recognition network for this feature,and puts forward a time stamp information constraint mechanism,which combines the text semantic constraint information with the optical characteristics to achieve the effect of identifying standard text,and enhance formatting specifications and numerical rationality of the output timestamp text.Classic text recognition algorithms based on optical features are completely surpassed in terms of full match rate and editing distance.