收稿日期:2022-10-04基于机器视觉的多种类标签识别方法研究朱传浩,欧阳八生(南华大学机械工程学院,湖南衡阳421001)摘要:针对目前印刷标签复杂多样难以识别和分类,以及各种缺陷造成识别准确率低的问题,提出了一种基于机器视觉的多种类标签检测方法。首先搭建由相机、镜头和光源组成的采样平台,将各种类标签采集后用作模型训练的数据集,然后利用最小外接矩形并稳健回归的方式对图像进行畸变矫正,通过Laplacian算子、高斯滤波算法、Otsu算法消除噪声产生的影响,最后建立了一个改进的CRNN+CTC网络结构模型,其中加入BN算法和Adam算法提高模型的泛化能力和收敛速度,使用双向BLSTM网络减小梯度消失或爆炸,再加入CTC损失函数实现输入数据与给定标签的对齐问题。实验结果表明,改进后的方法相较于传统分割字符算法,识别准确率提升至98.2%;相较于原CRNN+CTC算法,识别速度提升至37ms/张,达到了工业使用需求。关键词:多种类标签;稳健回归;机器视觉;CRNN+CTC中图分类号:TP391.9文献标志码:A文章编号:1009-9492(2023)02-0177-05ResearchonMulti-typeLabelRecognitionMethodBasedonMachineVisionZhuChuanhao,OuyangBasheng(SchoolofMechanicalEngineering,UniversityofSouthChina,Hengyang,Hunan421001,China)Abstract:Aimingattheproblemsofcomplexanddiverseprintedlabelsthataredifficulttoidentifyandclassify,aswellasthelowrecognitionaccuracycausedbyvariousdefects,amulti-typelabeldetectionmethodbasedonmachinevisionwasproposed.Firstly,asamplingplatformcomposedofcamera,lensandlightsourcewasbuilt,andvariouslabelswerecollectedandusedasdatasetsformodeltraining.Then,theimagedistortionwascorrectedbyusingtheminimumexternalrectangleandrobustregression,andtheinfluenceofnoisewaseliminatedbyLaplacianoperator,GaussianfilteringalgorithmandOtsualgorithm.Finally,animprovedCRNN+CTCnetworkstructuremodelwasestablished,inwhichBNalgorithmandAdamalgorithmwereaddedtoimprovethegeneralizationabilityandconvergencespeedofthemodel,bidirectionalBLSTMnetworkwasusedtoreducethegradientdisappearanceorexplosion,andCTClossfunctionwasaddedtorealizethealignmentproblembetweeninputdataandgivenlabels.Experimentalresultsshowthattherecognitionaccuracyoftheimprovedmethodis98.2%compare...