基于迁移学习和改进残差网络的棉花叶螨为害等级识别  

Identification of harm grades of cotton spider mites based on transfer learning and improved residual network

作　　者：张䶮 周保平[1] 王昱 冯洁 叶凡恺 何云龙 ZHANG Yan;ZHOU Baoping;WANG Yu;FENG Jie;YE Fankai;HE Yunlong(College of Information Engineering,Tarim University,Alar 843300,Xinjiang,China;Key Laboratory of Modern Agricultural Engineering,Tarim University,Alar 843300,Xinjiang,China)

机构地区：[1]塔里木大学信息工程学院,新疆阿拉尔843300 [2]塔里木大学现代农业工程重点实验室,新疆阿拉尔843300

出　　处：《浙江农业学报》2023年第7期1729-1739,共11页Acta Agriculturae Zhejiangensis

基　　金：国家自然科学基金(61563046)。

摘　　要：针对人工诊断棉叶螨害分级准确率低、耗时长、成本高的问题,提出一种基于迁移学习和改进残差网络的棉花叶螨为害等级识别方法。以3种受害等级的棉花叶片与健康叶片图像作为对象,分别于单一背景和自然环境下采集图像,构建图像数据集。首先,利用PlantVillage数据集预训练模型,使用数据增强技术对数据集进行数据增强,扩充训练样本;然后,在ResNet50网络模型的基础上,引入焦点损失函数,在不同网络层嵌入注意力机制模块,并加入Dropout正则化构建改进的ResNet50模型;最后,对比不同模型的识别效果。结果表明:同时在深层和浅层引入注意力机制模块,设定动量为0.9、学习率为0.001时,改进的ResNet50模型具有最好的分类效果,优于ResNet50、VGG16、MobileNet、AlexNet和SENet模型,对棉叶螨危害等级的平均识别准确率达到97.8%。In view of the low accuracy,long time consuming and high cost of traditional artificial diagnosis of cotton spider mites,a method for detection and classification of harm grade was proposed based on transfer learning and improved residual network.The cotton leaf images either healthy or with 3 harm grades of cotton spider mites were collected both in single background and natural environment to construct image dataset.First,the PlantVillage dataset was used to pretrain the model.Data augmentation was carried out to expand training samples.Then,based on the original ResNet50 network,an improved ResNet50 network was constructed by introducing focal loss functions,embedding attention mechanism modules in different network layers,and optimizing with the Dropout regularization.Finally,the performance of the improved ResNet50 network was compared with other models.It was shown that with the attention mechanism module introduced both in the deep and shallow layers,the momentum being 0.9,and the learning rate being 0.001,the improved ResNet50 network had the best classification effect,which was superior than the original ResNet50,VGG16,MobileNet,AlexNet and SENet models,with the average recognition accuracy of 97.8%.

关 键 词：棉花叶螨 受害等级 ResNet50网络 迁移学习 焦点损失函数 注意力机制 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]