·机械制造·杨涛，等·联合损失监督的高频工件深度学习识别算法基金项目:四川省重大科技专项(18ZDZX0140)第一作者简介:杨涛(1997—)，男，四川雅安人，硕士研究生，研究方向为图像处理与识别模式。DOI:10．19344/j．cnki．issn1671－5276．2023．01．007联合损失监督的高频工件深度学习识别算法杨涛1，欧阳1，苏欣2，吴学杰1，李柏林1(1．西南交通大学机械工程学院，四川成都610031;2．中国电子科技集团公司第十研究所，四川成都610036)摘要:针对高频工件种类多、类间相似度较高造成的识别准确率低的问题，提出一种联合损失监督的深度学习识别算法。搭建基于卷积神经网络的图像特征向量编码模型，采用角度余量损失替换SoftMax损失，以减小工件类内特征之间的距离，完成同类工件的鲁棒性表示;引入隔离损失以增大异类工件特征之间的距离，实现异类工件的良好性区分。实验结果表明:该方法相较于传统的图像识别方法，识别准确率更高;相较于单一的角度余量和隔离损失，识别准确率分别提高了3．97%和13．88%。关键词:工件识别;联合损失;监督学习;卷积神经网络中图分类号:TP391．41文献标志码:A文章编号:1671-5276(2023)01-0030-04HighFrequencyWorkpieceDeepLearningＲecognitionAlgorithmBasedonJointLossSupervisionYANGTao1，OUYang1，SUXin2，WUXuejie1，LIBailin1(1．SchoolofMechanicalEngineering，SouthwestJiaotongUniversity，Chengdu610031，China;2．The10thＲesearchInstituteofCETC，Chengdu610036，China)Abstract:Toimprovethelowrecognitionaccuracycausedbywidevarietiesofhighfrequencyartifactsandhighsimilaritybetweenclasses，adeeplearningalgorithmwithjointlosssupervisionisproposed．Animagefeaturevectorencodingmodelisbuiltbasedonconvolutionalneuralnetwork，andtheSoftMaxlossisreplacedbytheanglemarginlosstoreducethedistancebetweenthefeatureswithintheworkpiececlassandcompletetherobustrepresentationofsimilarworkpieces．Theisolationlossisintroducedtoincreasethedistancebetweenthefeaturesofheterogeneousworkpiecesandachievegooddiscriminationofheterogeneousworkpieces．Theexperimentalresultsshowthattherecognitionaccuracyoftheproposedmethodishigherthanthatofthetraditionalimagerecognitionmethod，withthesingleanglemarginincreasingby3．97%andisolationloss13．88%respectively．Keywo...
