复杂背景下苹果叶片病害严重程度自动估计  

Estimation of apple leaf disease severity under complex background based on semantic segmentation network

作　　者：范红宇 刘博[1,2] 么炜 程洪[1,2] FAN Hongyu;LIU Bo;YAO Wei;CHENG Hong(College of Information Science and Technology,Hebei Agricultural University,Baoding 071001,China;Hebei Key Laboratory of Agricultural Big Data,Baoding 071001,China)

机构地区：[1]河北农业大学信息科学与技术学院,河北保定071001 [2]河北省农业大数据重点实验室,河北保定071001

出　　处：《河北农业大学学报》2025年第2期88-100,共13页Journal of Hebei Agricultural University

基　　金：河北省省级科技计划(20327404D,21327404D);河北省省属高校基本科研业务费研究项目(KY2022072).

摘　　要：为解决苹果叶片病害严重程度难以在复杂环境下进行自动估计的问题,本文以苹果锈病、疮痂病、蛙眼叶斑病为研究对象,提出了一种基于两阶段语义分割模型的苹果叶片病害严重程度估计方法。第一阶段,针对自然环境下叶片分割精度不高的问题,对PSPNet中金字塔池化层进行优化并联加入了可变形卷积层(Deformable convolution,DCN),从复杂环境下分割苹果目标叶片。第二阶段,采用UNet网络分割目标叶片的病斑,在其骨干网络VGG16的各激活层前引入批归一化层防止过拟合;并采用双线性插值替换解码器中转置卷积进行上采样,避免转置卷积引起的棋盘效应,对第一阶段得到的叶片结果进行病斑分割,病斑与叶片的面积比即为估计的叶片病害程度。实验结果表明,本文提出的两阶段分割模型能够满足复杂环境叶片和病斑的分割需求,叶片的分割精度达到98.76%,病斑的分割精度达到99.69%。在病害严重程度估计方面,准确率、F1值均优于LD-Deeplabv3+、PUNet、DUNet等方法。此外,本文进一步分析3种病害的估计值与真实值的决定系数R^(2)与MRE,锈病、疮痂病、蛙眼叶斑病的R^(2)分别为0.9831、0.9707、0.9803,MRE分别为1.0953%、1.2401%、1.1086%。该研究可为自然环境下其他植物叶片的分割或病斑的检测等研究工作提供参考。In order to solve the problem that it is difficult to automatically estimate the severity of apple leaf diseases in complex environments,this paper proposed a method for estimating the severity of apple leaf diseases based on a two-phase semantic segmentation model,tncluding apple rust,scab,and frog's eye leaf spot disease as the research objects.In the first stage,the pyramid pooling layer in PSPNet was optimized and coupled with a deformable convolution layer(DCN)to segment the apple target leaves from the complex environment.In the second stage,the UNet network was used to segment the diseased spots of the target leaves,and a batch normalization layer was introduced in front of each activation layer of its backbone network VGG16 to prevent overfitting;and bilinear interpolation was used to replace the transposed convolution in the decoder for up-sampling to avoid the checkerboard effect caused by the transposed convolution,and the results of the leaves obtained in the first stage were segmented into diseased spots,and the area ratio of diseased spots to leaves was considered to estimate the degree of leaf disease.The experimental results showed that the two-stage segmentation model proposed in this paper met the segmentation needs of leaves and lesions in complex environments,and the segmentation accuracy of leaves reached 98.76%,and the segmentation accuracy of lesions reached 99.69%.The accuracy and F1-score of disease severity estimation were better than those of LD-Deeplabv3+,PUNet,DUNet and other methods.In addition,this paper further analyzed the decision system of R^(2)and MRE between the estimated and true values of the three diseases.The R^(2)of rust,scab,and frog’s eye leaf spot were 0.9831,0.9707,and 0.9803,respectively,and the MRE were 1.0953%,1.2401%,and 1.1086%,respectively.This study can provide a reference for research work such as segmentation of other plant leaves or detection of disease spots in natural environments.

关 键 词：苹果病害 病害严重程度估计 可变形卷积 语义分割 两阶段网络 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]