摘要
为了提高自然环境下苹果病虫害的识别准确率和识别效率,提出了具有动态学习特征的VGG-F苹果病虫害识别模型。首先,依据常见的苹果病害和虫害类型构建图像数据集,同时采用Retinex算法对数据集中的含雾图像进行增强处理;然后选择网络层数较少的VGG-F网络模型作为迁移学习对象,并依据数据样本特性对重训练过程进行学习率动态调整,以及基于试验对比选取最佳动量值;最后,利用数据集对三种不同模型进行重训练和识别效果对比测试。数值测试结果表明:相比于原始VGG-F模型和深层模型VGG-19,苹果病虫害识别准确率分别提升了5%和0.63%,且该模型的重训练时间最短,从而验证了苹果病虫害识别模型的有效性。
In order to improve the recognition accuracy and recognition efficiency of apple pests and diseases in natural environments,a VGG-F apple pests and diseases recognition model with dynamic learning characteristics was proposed.Firstly,the image dataset was constructed according to common apple diseases and pests.And the images containing fog inside the dataset were enhanced by Retinex algorithm.Then,the VGG-F network model with fewer network layers was selected as the object of transfer learning.According to the characteristics of the dataset,the learning rate was adjusted dynamically and the optimal momentum was selected by experimental comparison.Finally,the dataset was used for three different models to retrain and test recognition effect.The numerical test results show that,compared with the original VGG-F model and the deep model VGG-19,the recognition accuracy of the proposed model for apple pests and diseases is improved by 5%and 0.63%respectively,and the retraining time of the model is the shortest,which verifies the validity of the proposed model.
作者
于洪涛
袁明新
王琪
江亚峰
YU Hong-tao;YUAN Ming-xin;WANG Qi;JIANG Ya-feng(School of Mechanics and Power Engineering,Jiangsu University of Science and Technology,Zhangjiagang 215600,China;School of Computer Science and Technology,Tianjin University,Tianjin 300000,China)
出处
《科学技术与工程》
北大核心
2019年第32期249-253,共5页
Science Technology and Engineering
基金
国家重点研发计划重点专项(2016YFD0700903)资助
关键词
苹果
病虫害识别
VGG-F模型
迁移学习
动态学习率
动量
apple
diseases and pests recognition
VGG-F model
transfer learning
dynamic learning rate
momentum
