基于改进Faster R-CNN的苹果叶部病害识别方法  

Apple disease identification using improved Faster R-CNN

作　　者：王云露 吴杰芳[2] 兰鹏[1] 李凤迪 葛成恺 孙丰刚[1] WANG Yunlu;WU Jiefang;LAN Peng;LI Fengdi;GE Chengkai;SUN Fenggang(College of Information Science and Engineering,Shandong Agricultural University,Taian 271018,China;School of Mathematics and Statistics,Taishan University,Taian 271000,China)

机构地区：[1]山东农业大学信息科学与工程学院,泰安271018 [2]泰山学院数学与统计学院,泰安271000

出　　处：《林业工程学报》2022年第1期153-159,共7页Journal of Forestry Engineering

基　　金：山东省重大科技创新工程项目(2019JZZY010706);山东省重点研发计划项目(2017CXGC0206,2019GNC106106);山东省自然科学基金面上项目(ZR2019MF026)。

摘　　要：针对苹果叶片图像中小尺度病斑和复杂背景带来的病斑目标难以精确定位和识别的问题,以苹果的斑点落叶病、黑星病、灰斑病、雪松锈病和花叶病为研究对象,提出一种基于改进Faster R-CNN的苹果叶片病害识别方法。先通过数据增广操作对训练集数据进行扩充以增强模型鲁棒性,再通过对增广训练集图像进行训练来得到一个可靠的病害识别模型。改进后的模型使用拆分注意力网络(ResNest)作为骨干特征提取网络,使模型更加关注对提升病斑检测性能有用的信息,以增强模型对特征的提取能力;通过添加特征金字塔网络(FPN)进行多尺度特征融合,以增强特征信息的鲁棒性,提高模型的泛化能力;采用级联机制对建议框生成机制进行优化,使检测框定位更加准确。改进后的Faster R-CNN模型的平均精度均值(mAP)达到86.2%,与改进前相比,其平均精度提升了8.7%,对单张病害图像的识别准确率达到98.3%,单张图像平均检测时间0.092 s,能有效识别苹果叶片病斑。实验结果表明,改进后的Faster R-CNN模型能准确快速地实现对苹果叶片小目标病斑和复杂背景下病斑的识别,提升模型识别的精准度。该识别方法可在实际场景下使用,无须特意采摘叶片实现对苹果叶片病害的无损测量识别,可为苹果病害的早期干预和治疗提供科学依据。To analyze apple diseased leaf images,it is difficult to locate and identify these diseased leaves with small scale lesion and complex background in the actual application scenarios.In this study,five apple leaf diseases,i.e.,alternaria leaf spot,apple scab,gray spot,cedar rust and mosaic were investigated,and an improved Faster R-CNN based apple diseased leaf detection method was proposed.Firstly,the training set data was expanded through the data augmentation operation(including rotating,random brightness enhancement,random chromaticity enhancement,random contrast enhancement and sharpening)to enhance the robustness of the model.Then the augmented training set images were trained through the improved Faster R-CNN to make the detection model more reliable.For the improved Faster R-CNN model,the attention separation mechanism ResNest(split-attention networks)was adopted as the backbone to make the model focusing on the more useful information to enhance the feature extraction ability according to the feature representation through the weighted combination.To enhance the robustness of the feature information and improve the generalization ability,a feature pyramid network(FPN)was added to fuse multi-scale features,which effectively used the deep and shallow features of the network.Meanwhile,the cascade mechanism was adopted to optimize the generation mechanism of the suggestion box,so that the detection box location was more accurate.The mAP(mean average precision)of the improved model reached 86.2%,which is 8.7%higher than that of the previous Faster R-CNN model.The accuracy of the model reached 98.3%and the average detection time of the model was 0.092 s,which can effectively identify apple leaf lesions.The experimental results showed that the improved Faster R-CNN model could accurately and quickly identify small target lesions of apple leaves and lesions under complex background,and improve the accuracy of model recognition.The images in the data set included picked leaves and non-picked leaves,so this method

关 键 词：苹果病害识别 深度学习 Faster R-CNN ResNest 多尺度特征融合 级联机制 

分 类 号：S436.611[农业科学—农业昆虫与害虫防治] TP391.4[农业科学—植物保护]