一种改进Transformer的仪表字符识别算法  

An Improved Instrument Character Recognition Algorithm for Transformer

作　　者：寇文博 屈八一[1] 李智奇[2] KOU Wenbo;QU Bayi;LI Zhiqi(Chang'an University,Xi’an 710054,China;Xi’an University of Electronic Science and technology,Xi’an 710071,China)

机构地区：[1]长安大学,西安710054 [2]西安电子科技大学,西安710071

出　　处：《自动化与仪器仪表》2022年第7期284-288,共5页Automation & Instrumentation

基　　金：中央高校基本科研业务费专项资金(300102240107);陕西省自然科学基金资助项目(211424190465);国家自然科学基金资助项目(11873039)。

摘　　要：针对传统工业仪表在智能化改造过程中,非智能仪表不能输出数字信号的问题,以数显式流量计仪表字符为研究对象,以提高仪表字符识别准确度和识别速度为目的,采用一种改进Transformer的仪表字符识别算法。该算法添加修改的EfficientNetV2s模型作为特征提取层;在Transformer模型的多头注意力中使用多头加权融合机制;并增加相对位置编码。实验结果表明,改进得到的基准模型识别精度提高到99.6%,识别速度达到107.3 ms,达到了预期的识别效果。Aiming at the problem that nonintelligent instruments cannot output digital signals in the process of intelligent transformation of traditional industrial instruments,taking the instrument characters of digital display flowmeter as the research object,in order to improve the accuracy and speed of instrument character recognition,an improved transformer instrument character recognition algorithm is adopted.The modified efficientnetv2 s model is added as the feature extraction layer;The multi-head weighted fusion mechanism is used in the multi-head attention of transformer model;And add relative position coding.The experimental results show that the recognition accuracy of the improved benchmark model is improved to 99.6%,the recognition speed is 107.3 ms,and the expected recognition effect is achieved.

关 键 词：仪器仪表 计算机视觉 字符识别 TRANSFORMER 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP183[自动化与计算机技术—计算机科学与技术]