基于改进ShuffleNetV2的轻量化饲料原料种类识别模型  

A lightweight model for identifying types of feed raw material based on improved ShuffleNetV2

作　　者：田敏 牛智有[1] 刘梅英[1] TIAN Min;NIU Zhiyou;LIU Meiying(College of Engineering/Ministry of Agriculture and Rural Affairs Key Laboratory of Smart Farming for Agricultural Animals,Huazhong Agricultural University,Wuhan 430070,China)

机构地区：[1]华中农业大学工学院/农业农村部智慧养殖技术重点实验室,武汉430070

出　　处：《华中农业大学学报》2025年第2期105-115,共11页Journal of Huazhong Agricultural University

基　　金：国家自然科学基金项目(32072765)。

摘　　要：针对目前配合饲料加工过程中生产饲料原料入仓环节人工识别粉碎程度、颜色和形状相近的入仓原料时存在种类识别困难的问题,基于轻量化卷积神经网络模型ShuffleNetV2,提出一种识别精度更高、计算复杂度更小且适用于饲料原料图像种类识别的ShuffleNetV2-EH轻量化模型。首先在ShuffleNetV2网络模型结构中引入注意力机制ECA(efficient channel attention),根据输入自适应调整通道权重,提升网络模型对饲料原料图像重要特征的感知能力;其次将ReLU替换为HardSwish激活函数,在不增加额外的权重和偏置参数的前提下,提升模型的识别准确率;最后在保证模型识别精度的基础上,对ShuffleNetV2网络模型结构进行调整,减少模型的参数量以及计算量。结果显示,ShuffleNetV2-EH模型在8种饲料原料图像测试集上的识别准确率为99.13%,与原ShuffleNetV2模型相比提升1.38百分点,其精确率、召回率和F1分数分别提升1.45、1.63和1.62百分点,模型参数量和浮点运算量较之前分别减少352 092个和45.27×10^(6);且综合性能优于经典卷积神经网络模型AlexNet、VggNet16、GoogLeNet和ResNet18。结果表明,改进后的ShuffleNetV2模型较好地平衡了模型的计算复杂度和识别精度,为入仓环节的饲料原料在线识别提供了算法基础。A lightweight model of ShuffleNetV2-EH with higher accuracy of identification,lower complexity of computation,and suitable for identifying the types of feed raw material based on the lightweight convolutional neural network model ShuffleNetV2 to achieve rapid identification of warehousing feed raw materials and solve the difficulties in manually identifying the types of feed raw materials with similar crushing degree,color,and shape in currently processing and producing the combined feed raw materials.Firstly,the efficient channel attention(ECA) mechanism was introduced into the structure of ShuffleNetV2 network model,which adaptively adjusts channel weights based on input to enhance the ability of network model to percept important features in images of feed raw materials.Secondly,ReLU was replaced with HardSwish activation function to improve the recognition accuracy of the model without adding additional weights and parameters of bias.Finally,the structure of ShuffleNetV2 network model was adjusted to reduce the number of parameters and the complexity of computation in the model on the basis of ensuring the recognition accuracy of model.The results showed that the recognition accuracy of ShuffleNetV2-EH model on image sets from 8 types of feed raw materials was 99.13%,1.38% higher than that of the original ShuffleNetV2 model.Its accuracy,recall,and F1 score increased by 1.45%,1.63%,and 1.62 %,respectively.The number of model parameters and floating-point operations decreased by 352 092and 45.27×10^(6),compared to that of the original model.The overall performance was superior to classical convolutional neural network models including AlexNet,VggNet16,GoogLeNet,and ResNet18.It is indicated that the improved ShuffleNetV2 model well balances the complexity of computation and the recognition accuracy of the model,providing an algorithm foundation for online identification of feed raw materials in the warehousing process.

关 键 词：饲料原料 种类识别 轻量化模型 ShuffleNetV2 注意力机制 

分 类 号：S512.2[农业科学—作物学] TP183[自动化与计算机技术—控制理论与控制工程]