基于注意力机制及多尺度特征融合的番茄叶片缺素图像分类方法  

Image classification method for tomato leaf deficient nutrient elements based on attention mechanism and multi-scale feature fusion

作　　者：韩旭 赵春江 吴华瑞[2,3] 朱华吉[2,3] 张燕 Han Xu;Zhao Chunjiang;Wu Huarui;Zhu Huaji;Zhang Yan(College of Information Engineering,Northwest A&F University,Yangling 712100,China;National Engineering Research Center for Information Technology In Agriculture,Beijing 100097,China;Beijing Research Center for Information Technology In Agriculture,Beijing 100097,China)

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

出　　处：《农业工程学报》2021年第17期177-188,共12页Transactions of the Chinese Society of Agricultural Engineering

基　　金：国家重点研发计划资助(2020YFD1100602);国家自然科学基金资助(61871041);财政部和农业农村部:国家现代农业产业技术体系资助(CARS-23-C06)

摘　　要：针对番茄早期缺素性状不明显及各生长期特征差异较大所导致的特征区域尺寸不一致、难提取、难辩别等问题,提出了一种基于注意力机制及多尺度特征融合卷积神经网络(Multi-Scale Feature Fusion Convolutional Neural Networks based on Attention Mechanism,MSFF-AM-CNNs)的番茄叶片缺素图像分类方法。首先根据番茄叶片缺素特点提出了多尺度特征融合结构(Multi-Scale Feature Fusion Module,MSFF Module);其次在DenseNet基础上,结合浅层网络主要提取纹理、细节特征,深层网络主要提取轮廓、形状特征的特点分别提出具有针对性的特征提取方法,通过不同形式引入注意力机制及多尺度特征融合结构,使全局多尺度信息融合多个特征通道、选择性地强调信息特征并达到对特征精准定位的功能;同时引入Focal Loss函数以减少易分类样本的权重。试验结果表明,MSFF-AM-CNNs的平均召回率、平均F1得分、平均准确率较原模型DenseNet-121均大幅提升,其中缺氮和缺钾叶片的准确率分别提高了8.06和6.14个百分点,召回率分别提高了6.31和5.00个百分点,F1得分分别提高了7.25和5.55个百分点,平均识别准确率可达95.92%,具有较高的识别准确率及广泛的适用性,能够满足番茄叶片缺素图像的高精度分类需求,可为植物叶片缺素识别提供参考。This study aims to realize the accurate identification of deficient nutrients elements in tomato leaves.An experiment was conducted on the lack of nutrients in the climate chamber in the laboratory of Big Data Intelligence Department of the Beijing Academy of Agriculture and Forestry Sciences,China.An artificial climate chamber was also selected to regulate the growth environment factors of tomato plants for the specific lack of nutrients.Three types of nutrient deficiency groups were set,namely nitrogen deficiency,phosphorus deficiency,and potassium deficiency,as well as a normal control group.The experiment was started with the appearance of nutrient-deficient traits in seedlings,and the images of nutrient-deficient leaves were then collected according to the growth stages.The experimental results show that there were diversity and differences in the traits of tomato nutrient deficiency.Specifically,there were relatively small changes of leaves in the early stage of tomato nutrient deficiency.Furthermore,it was difficult to capture the details and textures,due to the smaller area of traits.For example,the manifestation of phosphorus deficiency was that the leaves gradually turn purple along the veins.The trait details were hardly identified in the early stage of phosphorus deficiency,due mainly to the mostly small vein structure.Particularly,tomato leaves under different conditions of nutrient deficiency presented similar color and texture characteristics at a certain stage.For example,the leaves were both slightly yellow in the early stage of nitrogen deficiency and the early stage of potassium deficiency.The only slight difference was the characteristic display of morphology in the size of characteristic areas.There were obvious differences in color and texture at different stages under the same nutrient deficiency.The images were collected from the climate chamber to serve as the experimental data.An attempt was made on the inconsistency of feature area size,and the difficulty of feature extraction,resulting

关 键 词：图像识别 缺素 蔬菜 多尺度特征融合 注意力机制 卷积神经网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]