摘要
[目的]小麦秸秆表皮结构完整性是判断其资源化利用效果的关键因素之一,目前尚难以对秸秆破碎程度进行量化分析,为了实现小麦秸秆表皮结构完整性的分类,本研究提出了一种改进的GhostNet轻量级卷积神经网络,用于对小麦秸秆表皮显微图像进行完整性分类。[方法]基于小麦秸秆表皮显微成像技术,将迁移学习引入GhostNet中,降低了模型过拟合的风险,同时采用了Dropout层以提升网络的分类准确率。为了验证该方法的有效性,利用4320幅小麦秸秆表皮显微图像进行训练和验证,同时与ShuffleNet V2、ResNet 50和AlexNet深度学习网络进行了对比。[结果]试验结果表明,改进的GhostNet网络模型的分类准确率为99.2%,分别比ShuffleNet V2、ResNet 50和AlexNet提高了14.55%、3.66%和3.44%,为了验证该模型的鲁棒性,分别对高斯噪声和不同亮暗程度影响进行了测试,测试结果表明,改进后的GhostNet网络模型依然可以取得最佳的分类效果。[结论]该方法应用于小麦秸秆表皮显微图像的完整性分类是有效的、可行的,该方法可为秸秆预处理技术效率的量化分析提供参考。
[Objectives]The integrity of wheat straw epidermal is one of the key factors to judge its feasibility of resource utilization.At present,it is difficult to quantify the degree of wheat straw epidermal fragmentation.In order to classify the integrity of wheat straw epidermis,based on the micro-imaging technology of wheat straw epidermis,an improved lightweight GhostNet convolutional neural network was proposed to classify the intact and broken epidermis of wheat straw.[Methods]Transfer learning was used in this research,in order to reduce the risk of model’s over fitting,meanwhile,Dropout layer was adopted to improve the accuracy of the model.In order to verify the effectiveness of this model,4320 micrographs of wheat straw were trained and verified,and compared with ShuffleNet V2,ResNet 50 and AlexNet deep learning network models.[Results]The results showed that the classification accuracy of the improved GhostNet was 99.2%,which was 14.55%,3.66%and 3.44%higher than that of ShuffleNet V2,ResNet 50 and AlexNet,respectively.In order to verify the robustness of these models,the effects of Gaussian noise and the brightness of micrographs were tested respectively.The test results showed that the GhostNet model still had the highest accuracy of classification.[Conclusions]All the results indicated that the proposed method was effective and feasible in the classification of wheat straw epidermis integrity.This method can provide necessary technical support for the quantitative analysis of the pretreatment efficiency of wheat straw.
作者
张倩如
王云飞
吕帅朝
宋磊
尚钰莹
宋怀波
ZHANG Qianru;WANG Yunfei;Lü Shuaichao;SONG Lei;SHANG Yuying;SONG Huaibo(College of Mechanical and Electronic Engineering/Key Lab of Agricultural Internet of Things,Ministry of Agriculture and Rural Affairs,Northwest A&F University,Yangling 712100,China)
出处
《南京农业大学学报》
CAS
CSCD
北大核心
2022年第4期788-798,共11页
Journal of Nanjing Agricultural University
基金
国家重点研发计划项目(2019YFD1002401)。
