改进注意力机制的多叉树网络多作物早期病害识别方法  

Multi-tree Network Multi-crop Early Disease Recognition Method Based on Improved Attention Mechanism

作　　者：高荣华 白强[1,2,3] 王荣 吴华瑞 孙想[1,2] GAO Rong-hua;BAI Qiang;WANG Rong;WU Hua-rui;SUN Xiang(Research Center of Information Technology,Beijing Academy of Agriculture and Forestry Sciences,Beijing 100097,China;National Engineering Research Center for Information Technology in Agriculture,Beijing 100097,China;College of Information Engineering,Northwest A&F University,Xianyang,Shaanxi 712100,China)

机构地区：[1]北京市农林科学院信息技术研究中心,北京100097 [2]国家农业信息化工程技术研究中心,北京100097 [3]西北农林科技大学信息工程学院,陕西咸阳712100

出　　处：《计算机科学》2022年第S01期363-369,共7页Computer Science

基　　金：国家自然科学基金面上项目(61771058);北京市科技计划课题(Z191100004019007)。

摘　　要：在作物染病早期,及时获取作物病害信息,判别染病原因和严重程度,从而对症下药,能够及时防治病害扩散造成的作物产量下降。针对传统深度学习网络对作物早期病害识别方法准确率低的问题,基于病害特征图像各通道包含的信息量不同,及多层感知机(Multilayer Perceptron,MLP)能逼近任意函数的特点,提出了一种改进注意力机制的多叉树网络作物早期病害识别方法,将注意力机制融合残差网络对病害特征重校准(SMLP_Res);同时结合多叉树结构对具有较高特征提取能力的SMLP_ResNet(Squeeze-Multi-Layer Perceptron ResNet)网络进行扩展,构建的多叉树SMLP_ResNet网络模型可以简化多作物早期病害识别任务,有效提取早期病害特征。实验中使用Plant Village和AI Challenger 2018两种数据集对18层的ResNet,SE_ResNet,SMLP_ResNet这3种网络模型,以及同等结构的多叉树结构模型进行训练,验证了SMLP_Res和多叉树结构对作物病害识别模型的影响。通过实验分析得到18层的ResNet,SE_ResNet,SMLP_ResNet这3种网络模型在病害特征较明显的Plant Village数据集上病害识别的准确率均达到99%以上,但在早期病害数据集AI Challenger 2018上的准确率均不超过87%,SMLP_ResNet因加入了SMLP_Res模块,故对作物病害数据特征提取较为充分,检测结果较好。多叉树结构的3种早期病害识别模型,在AI Challenger 2018数据集上识别准确率均有明显提升,多叉树SMLP_ResNe较其余两种模型具有较好的性能,其中樱桃早期病害识别准确率为99.13%,检测结果最佳。文中提出的多叉树SMLP_ResNet作物早期病害识别模型,能够简化识别任务,抑制噪声传输,达到了较高的准确率。In the early stage of crop infection,timely acquisition of crop disease information,identification of the cause and severity of disease,and the right remedy,can prevent and control the decline in crop yield caused by the spread of the disease in time.In view of the low accuracy of traditional deep learning network for early crop disease recognition,based on the difference in information contained in each channel of the disease feature image and the characteristics of multilayer prceptron(MLP)that can approximate any function,a multi-tree network crop early disease identification method based on improved attention mechanism is proposed.It combines the attention mechanism with residual network to recalibrate disease features(SMLP_Res).At the same time,combined with the multi-tree structure,the SMLP_ResNet network with high feature extraction ability is expanded,and the constructed multi-tree SMLP_ResNet network model can simplify the task of early disease recognition of multiple crops and effectively extract early disease features.In experiments,Plant Village and AI Challenger 2018 are used to train18-layer model ResNet,SE_ResNet,SMLP_ResNet,as well as the multi-tree structure model with the same structure,to verify the influence of SMLP_Res and multi-tree structure on crop disease recognition models.Experimental analysis shows that,the disease recognition accuracy of the three network models on Plant Village dataset with obvious disease features is more than 99%,but their accuracy on the early disease data set AI Challenger 2018 is not more than 87%.SMLP_ResNet has sufficient feature extraction of crop disease data due to the addition of SMLP_Res module,and the detection results are better.The three early disease recognition models of the multi-tree structure significantly improves the recognition accuracy on AI Challenger 2018 dataset.The multi-tree SMLP_ResNe has better performance than the other two models,and the early disease recognition accuracy of cherry is 99.13%,the detection result is the best.The propos

关 键 词：早期病害识别 多叉树 损失函数 注意力机制 残差网络 

分 类 号：S431[农业科学—农业昆虫与害虫防治] TP391[农业科学—植物保护]