鄢江苏省科技计划项目渊BY2022134冤3D视觉技术在汽车轮胎字符识别中的应用鄢顾涛罗印升宋伟渊江苏理工学院袁江苏常州213001冤Applicationof3DVisionTechnologyinAutomobileTireCharacterRecognition摘要院基于2D视觉的轮胎字符识别方法存在效率低尧精度差尧易受光照条件影响等不足袁从而导致系统工作不稳定遥因此袁提出了一种基于3D视觉的轮胎字符识别算法遥首先选用3D线扫激光传感器获取轮胎胎面字符的三维点云数据袁根据点云数据在Z轴上的高度特征将其转换为灰度值袁然后采用以ResNet50为骨干网络的改进型DBNet算法袁结合Nadam方法对DBnet算法进行训练优化遥在此基础上袁进一步采用模型剪枝技术袁在保证算法精度的同时袁压缩模型参数袁提升算法速度袁大幅减少了计算量遥结果表明袁在相关的4种检测算法下袁该方法获取到的数据集的准确率高于传统方式数据集8.05%~12.2%袁改进的DBNet检测算法结合CRNN识别算法后袁在该方法获取的数据集中预测准确率达到了95.45%袁单张图像预测速度由107ms缩减到了45ms袁模型大小也由142.7MB减少到了15.83MB袁为轮胎字符快速准确识别提供了一种新型的技术方案遥关键词院字符检识别曰改进DBNet曰深度学习曰点云预处理Abstract:Thetirecharacterrecognitionmethodbasedon2Dvisionhassomeshortcomings,suchaslowefficiency,pooraccuracyandeasytobeaffectedbyilluminationconditions,whichleadstothesystem'sunstablework.Therefore,atirecharacterrecognitionalgorithmbasedon3Dvisionisproposedinthispaper.Firstly,3Dlinesweeplasersensorisusedtoobtainthethree-dimensionalpointclouddataoftiretreadcharacters,whichisconvertedintograyvalueaccordingtotheheightcharacteristicsofthepointclouddataontheZ-axis.Then,theimprovedDBNetalgorithmwithResNet50asthebackbonenetworkisusedtotrainandoptimizetheDBnetalgorithmcombinedwithNadammethod.Onthisbasis,themodelpruningtechnologyisfurtherusedtocompressthemodelparameters,improvethealgorithmspeedandgreatlyre鄄ducetheamountofcalculationwhileensuringtheaccuracyofthealgorithm.Theresultsshowthat,underthefourrelateddetectionalgorithms,theaccuracyofthedatasetobtainedbytheproposedmethodis8.05%~12.2%higherthanthatofthetraditionalmethod.Thepredictionspeedofsingleimageisreducedfrom107msto45ms,andthemodelsizeisalsoreducedfrom142....
