基于图像分类的多类型数字仪表自动读取方法  

Automatic Reading Method for Multi Type Digital Instruments Based on Image Classification

作　　者：李冰冰 朱格 曹晗 李峰[1] 潘雨青[1] LI Bingbing;ZHU Ge;CAO Han;LI Feng;PAN Yuqing(School of Computer Science and Communication Engineering,Jiangsu University,Zhenjiang Jiangsu 212013)

机构地区：[1]江苏大学计算机科学与信息工程学院,江苏镇江212013

出　　处：《软件》2023年第12期70-75,共6页Software

摘　　要：为解决现有数字式仪表自动读取方法依赖于大样本训练且精度不高的问题,提出了两级识别机制,即先对仪表图像进行分类获取先验知识,然后基于先验知识获取表盘中感兴趣的区域并进行字符识别。该方法联合使用级联小波变换模块以及ELU激活函数改进了ResNet34网络模型,提高了仪表盘分类的准确性,并采用先验知识获取感兴趣区域,进一步避免了无关信息的干扰,减少了样本量。实验数据表明,文中方法是可行有效的,300张数字式仪表图像读取准确率为95.67%,相比于传统识别方法,该方法识别的数字式仪表的类型更多,效果更好。To solve the problem of existing digital instrument automatic reading methods relying on large sample training and low accuracy,a two-level recognition mechanism is proposed,which first classifies the instrument image to obtain prior knowledge,and then obtains the region of interest in the dial based on prior knowledge and performs character recognition.This method combines cascaded wavelet transform modules and ELU activation functions to improve the ResNet34 network model and improve the accuracy of dashboard classification,and adopting prior knowledge to obtain regions of interest further avoids interference from irrelevant information and reduces sample size.The experimental data shows that the method proposed in the article is feasible and effective,with an accuracy rate of 95.67%for reading 300 digital instrument images.Compared to traditional recognition methods,this method recognizes more types of digital instruments and has better results.

关 键 词：图像分类 ResNet34 先验知识 数字式仪表 自动读表 

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]