摘要
目的:解决食用菌种类识别传统方法靠人眼判断的不足,降低同一科属性状相似的识别出错率。方法:提出一种基于卷积神经网络的EfficientNet食用菌图像分类模型,利用不同设备和拍摄环境采集食用菌图像并建立数据集,通过模型训练技巧和网络技巧对模型性能进行提升,提出一种YWeight权重衰减方法来控制有效学习率,通过控制cross-boundary来影响模型的泛化性能。结果:该方法在自建数据集YMushroom上使EfficientNet-B0获得79.82%(+0.85%)top-1精度,默认训练过程仅获得78.97%。在公开数据集Fungus上使EfficientNet-B0达到87.62%(+0.78%)准确率,原始训练准确率为86.84%。结论:通过调整超参数使模型可接近最优解,通过权重衰减提升了食用菌图像分类模型的性能。
Objective:In order to solve the problem of edible fungus species identification,an EfficientNet edible fungus image classification model based on convolution neural network is proposed.Methods:Firstly,the edible fungus images were collected and the datasets were made according to different equipment and shooting environment,and then the model performance was improved through model training skills and network skills.A YWeight weight attenuation method was proposed to control the effective learning rate,and the generalization performance of the model was affected by controlling the cross-boundary.Results:This method makes EfficientNet-B0 obtain 79.82%(+0.85%)top-1 accuracy on the self built dataset YMushroom,and only 78.97%in the default training process.On the public dataset fungus,the accuracy of EfficientNet-B0 was 87.62%(+0.78%)and the original training accuracy was 86.84%.Conclusion:Experiments show that by adjusting the super parameters,the model finds a near optimal solution,and improves the performance of the edible fungus image classification model through weight attenuation,which provides a basis for the automatic management of edible fungus planting base in the future.
作者
姚芷馨
张太红
赵昀杰
YAO Zhi-xin;ZHANG Tai-hong;ZHAO Yun-jie(College of Computer and Information Engineering,Xinjiang Agricultural University,Urumqi,Xinjiang 830052,China)
出处
《食品与机械》
北大核心
2022年第11期117-124,共8页
Food and Machinery
基金
新疆维吾尔自治区重大科技专项(编号:2017A01002)。
