第60卷第2期/2023年1月/激光与光电子学进展0215002-1研究论文基于注意力机制的污水微型动物识别方法肖蕾*，蓝宗苗广东技术师范大学自动化学院，广东广州510665摘要为了精准掌握污水处理系统活性污泥中微型动物的种类，及时调整污水处理工艺，针对传统机器学习需要人工设计特征、提取特征、设计分类器等复杂过程的弊端，提出一种基于注意力机制和迁移学习相结合的污水活性污泥中微型动物的深度学习识别方法。在迁移学习的基础上，通过对传统的VGG16模型添加注意力模块（SE-Netblock），调整输出模块，采用数据增强方法扩充数据集。实验结果表明：相比于改进前的模型，改进后的模型（T-SE-VGG16）能够准确识别不同类型污水活性污泥中的微型动物，测试准确率为98.21%，提高了识别精度，缩短了训练时间，模型收敛速度快，泛化能力强。结果证实了T-SE-VGG16模型对污水活性污泥中的微型动物识别的可行性和可靠性。关键词深度学习；迁移学习；注意力机制；活性污泥；微型动物识别中图分类号TP389文献标志码ADOI：10.3788/LOP212628IdentificationofSewageMicroorganismsUsingAttentionMechanismXiaoLei*,LanZongmiaoCollegeofAutomation,GuangdongPolytechnicNormalUniversity,Guangzhou510665,Guangdong,ChinaAbstractToaccuratelyidentifymicroorganismspeciesintheactivatedsludgeofsewagetreatmentsystemsandmodifythewastewatertreatmentprocessinreal-time,usingtraditionalmachinelearningmethodsisachallengebecauseofvariouscomplicatedprocesses.Inthisstudy,adeeplearningapproachbasedontheintegrationofattentionmechanismandtransferlearningisproposedtoaccuratelyidentifythespeciesofmicroorganismsinsewage-activatedsludgebyovercomingtherequirementsofdevelopingfeaturesmanually,extractingfeatures,designingclassifiers,andothercomplicatedprocesses.Onthebasisoftransferlearning,theconventionalVGG16modelisenhancedbyincludingtheattentionmodule(SE-Netblock)andmodifyingtheoutputmodule,andthedatasetisexpandedusingthedataimprovementapproach.Experimentalfindingsdemonstratethatcomparedwiththemodelbeforetheenhancement,theenhancedmodel(T-SE-VGG16)canaccuratelyrecognizemicroorganismsinvarioustypesofsewage-activatedsludgewithatestaccuracyof98.21%,whichenhancestherecognitionaccuracyandreducesthetrai...
