JournalofJilinUniversitCInformationScienceEditionMay20232023年5月No.3Vol.41第41卷第3期吉林大学（信息科学版）文章编号：1671-5896(2023)03-0450-09基于改进ShuffleNetV2网络的岩石图像识别袁硕，刘玉敏²，安志伟'，王硕昌"，魏海军1（1.东北石油大学电气信息工程学院，黑龙江大庆163318；2.重庆科技学院电气工程学院，重庆401331）摘要：由于基于传统深度学习的岩石图像识别算法模型比较繁琐，而且应用于移动终端等需要一定的计算能力，因此很难实现对岩石类型的实时准确判别。为此，以ShuffleNetV2网络为基础，插入通道连接注意力机制ECA（EfficientChannelAttention）模块，使用Mish激活函数代替ReLU激活函数并引人轻量级网络部件中的深度可分离卷积。将该方法用于岩石图像识别，实验结果表明，改进后的算法结构简单，同时具有轻量化的特点，其识别精度达到94.74%，可在移动终端等有限资源环境下应用。关键词：岩石图像；有效通道注意力机制；Mish激活函数；ShuffleNet网络中图分类号：TP312文献标志码：ARockImageRecognitionBasedonImprovedShuffleNetV2NetworkYUANShuo',LIUYumin’,ANZhiwei',WANGShuochang',WEIHaijun'(1.SchoolofElectricalandInformationEngineering,NortheastPetroleumUniversity,Daqing163318,China;2.SchoolofElectricalEngineering,ChongqingUniversityofScienceandTechnology,Chongqing401331,China)Abstract:Therockimagerecognitionalgorithmmodelbasedontraditionaldeeplearningiscumbersomeandrequirescertaincomputingpowerwhenitisappliedtomobileterminals,soitisdifficulttorealizereal-timeandaccurateidentificationofrocktypes.BasedontheShuffleNetV2network,weinserttheECA(EfficientChannelAttention)moduleofthechannelconnectionattentionmechanism,usetheMishactivationfunctiontoreplacetheReLUactivationfunction,andintroducethedepthwiseseparableconvolutioninthelightweightnetworkcomponents.Experimentsareperformedonrockimageswiththismethod.Experimentsshowthattherecognitionaccuracyofthealgorithmreaches94.74%.Theimprovedalgorithmstructureisnotcomplexandmaintainsthecharacteristicsoflightweight,whichlaysafoundationforitsapplicationinlimitedresourceenvironmentssuchasmobileterminals.Keywords:rockimage;efficientchannelattention(ECA);Mishactivationfunction...
