摘要
为了降低指针式压力表的误读率,减轻人工读数压力,提高仪表读数的精度,设计了一种基于深度学习的指针式压力表读数方法。通过在DBNet网络结构基础上增加主干网络ResNet-18各个卷积层的通道数来提高模型的鲁棒性,重新设计了更适应指针式压力表刻度值检测的损失函数,在刻度值精准检测识别的基础上设计了极坐标展开的方法,将弧形的刻度值展开成一条直线,提高了读数的准确率。实验结果表明,最大误差仅1.05%,平均误差仅0.725%。相较于常用的Hough直线检测与ORB结合或DBNet+CRNN检测的方法,读数识别的平均误差大幅降低,为指针式压力表的自动读数提供了新的思路。
A pointer pressure gauge reading method based on deep learning is designed to reduce the misreading rate,alleviate the burden of manual reading,and improve the accuracy of instrument readings.By increasing the number of channels in each convolutional layer of the backbone network ResNet-18 based on the DBNet network structure to improve the robustness of the model,a loss function more suitable for the detection of pointer pressure gauge scale values is redesigned.On the basis of accurate detection and recognition of scale values,a polar coordinate expansion method is designed to expand the curved scale values into a straight line,so as to improve the accuracy of readings.The experimental results show that the maximum error of the reading is only 1.05%,and its average error is only 0.725%.In comparison with the commonly used methods of Hough line detection combined with ORB or DBNet+CRNN detection,the average error of its reading recognition is reduced significantly,which provides a new idea for the automatic reading of pointer pressure gauges.
作者
林鸿正
张斌
赵成龙
戴杰
湛敏
LIN Hongzheng;ZHANG Bin;ZHAO Chengong;DAI Jie;ZHAN Min(College of Metrology and Measurement Engineering,China Jiliang University,Hangzhou 310018,China;Hangzhou Laiting Technology Co.,Ltd.,Hangzhou 310018,China)
出处
《现代电子技术》
北大核心
2024年第7期165-169,共5页
Modern Electronics Technique
关键词
深度学习
指针式压力表
极坐标展开
自动读数
卷积循环神经网络
读数识别
deep learning
pointer pressure gauge
polar coordinate expansion
automatic reading
convolutional recurrent neural network
reading recognition
