基于语义分割和可见光谱图的作物叶部病斑分割方法  

Segmentation Method for Crop Leaf Spot Based on Semantic Segmentationand Visible Spectral Images

作　　者：李凯雨 张慧[2] 马浚诚 张领先[1] LI Kai-yu;ZHANG Hui;MA Jun-cheng;ZHANG Ling-xian(College of Information and Electrical Engineering,China Agricultural University,Beijing 100083,China;College of Information and Management Science,Henan Agricultural University,Zhengzhou 450018,China;Institute of Environment and Sustainable Development in Agriculture,Chinese Academy of Agricultural Sciences,Beijing 100081,China)

机构地区：[1]中国农业大学信息与电气工程学院,北京100083 [2]河南农业大学信息与管理科学学院,河南郑州450018 [3]中国农业科学院农业环境与可持续发展研究所,北京100081

出　　处：《光谱学与光谱分析》2023年第4期1248-1253,共6页Spectroscopy and Spectral Analysis

基　　金：国家自然科学基金项目(62176212)资助。

摘　　要：病害严重影响作物品质,并造成经济损失。病斑分割是病害定量诊断的重要过程,其分割结果可为后续的识别和严重度估算提供有效依据。由于病斑具有不规则性和复杂性,且自然环境下病斑可见光谱图像易受光照变化等影响,传统的图像处理方法对病斑图像分割存在准确率低、普适性低和鲁棒性不高等问题。该工作提出了基于语义分割和可见光谱图像的作物叶部病害病斑分割方法。首先,以花生褐斑病、烟草赤星病为研究对象,使用尼康D300s单反相机共采集到165张可见光谱图像。通过Matlab Image Labeler APP对病害可见光谱图像进行像素标记,分别标记出褐斑病病斑、赤星病病斑和背景区域。其次,对标记后的数据采用水平翻转、垂直翻转、改变亮度等图像扩充方式,获得1850份增强后样本数据集。为了节约计算成本,将数据集的像素分辨率调整为300×300。最后,基于FCN,SegNet和U-Net 3种语义分割网络,构建4种作物叶部病害病斑分割模型,探索了数据增强、病害类别对病斑分割模型的影响,并采用4种分割指标评价模型效果。结果表明:仅对于病斑分割,图像增强能够提高模型的分割精度,增强后FCN模型的平均精度(MP)和平均交并比(MIoU)分别为95.71%和93.36%。4个语义分割模型显著优于支持向量机(SVM),其中FCN与U-Net,SegNet-2和SegNet-4分割模型相比,能够有效避免光线变化等影响,病斑分割精度(P)和交并比(IoU)分别达到99.25%和97.55%。对于病斑分类分割实验,FCN对两种病害的分割精度Pd分别达到97.54%和90.41%,对两种病害的交并比IOUd分别为95.61%和70.30%,均优于其他3种分割模型。FCN能够在分割病斑的同时也准确地识别病害类别,有较好的泛化性和鲁棒性,实现了自然场景下作物叶部病害病斑的识别与分割,为计算混合病害严重度提供了技术参考。Diseases affect crop quality seriously and cause economic losses.Disease spot segmentation is an important process of identification and disease severity estimation,whose segmentation results can provide an effective basis for subsequent identification and severity estimation.Due to the irregularity and complexity of lesions,and the visible spectrum image of lesions in the natural environment is susceptible to be change in illumination,traditional image processing methods have low accuracy,low universality and robustness for image segmentation of lesions.In this regard,this article proposed a method for the segmentation of crop leaf diseases based on semantic segmentation and visible spectrum images.Firstly,taking peanut brown spot and tobacco brown spot as the research objects,165visible spectrum images were collected using a Nikon D300sSLR camera.The visible spectrum images of the diseases were pixel-labeled through the Matlab Image Labeler APP,and the brown spot and background area were respectively marked.Secondly,the labeled dataset adopted image enhancement methods such as horizontal flipping,vertical flipping,changing brightness,etc.,to obtain 1850enhanced sample data sets and randomly divided them into the training set,validation set and test set according to the ratio of 8∶1∶1.At the same time,in order to save computational cost,the pixel resolution of the data set was adjusted to 300×300.Finally,four types of disease spot segmentation models were constructed based on the three semantic segmentation networks of FCN,SegNet and U-Net.The effects of data enhancement and disease types on the lesion segmentation model were explored.Four segmentation indicators were used to evaluate the model’s segmentation effect.The test results showed that only for lesion segmentation,image enhancement could improve the segmentation accuracy of the model.The model’s Mean Precision(MP)and Mean Intersection over Union(MIoU)were 95.71%and 93.36%,respectively.The 4semantic segmentation models were significantly better

关 键 词：可见光光谱 作物 病斑分割 语义分割 全卷积神经网络 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]