基于半监督主动学习的小麦叶片病虫害区域分割方法  

Regional segmentation method for wheat leaf diseases and insect pests based on semi-supervised active learning

作　　者：安瑞钰 郝志斌 An Ruiyu;Hao Zhibin(College of Computer and Information Engineering,Tianjin Agricultural University,Tianjin 300392,China)

机构地区：[1]天津农学院计算机与信息工程学院,天津300392

出　　处：《天津农学院学报》2024年第2期87-93,共7页Journal of Tianjin Agricultural University

基　　金：国家自然科学基金项目(31700642)。

摘　　要：小麦叶片病虫害区域的准确分割对实现病虫害检测自动化与智能化有重要作用,对提高农作物经济效益有重要的理论价值和现实意义。为解决现有研究中样本标注工作量大、分类效果差等问题,本研究结合人工干预标注的主动学习策略,提出一种基于半监督主动学习的小麦叶片病虫害区域分割方法,通过提取边界特征、颜色空间特征和纹理特征,组成特征向量,使用直推式支持向量机作为分类器,在每一轮训练中,结合k-means算法随机生成待标记样本,通过人机交互系统干预样本标签提高分类效率。此外,本文构建了在新的病虫害区域分割方法基础上的对病虫害进行分类的LM神经网络模型。在图像分割实验中所提出模型与经典的小样本SVM模型以及半监督TSVM模型对比,分割效果排名第一,证明了所提出分割策略的优越性。在对病虫害识别实验中,模型验证结果表明采用半监督主动学习方法的LM神经网络分类模型得到的识别准确率为93.75%,具有良好的分类效果,能够为后续的病虫害防治提供有效依据。The accurate segmentation of wheat leaf disease and pest areas plays an important role in realizing the automation and intelligence of disease detection,and has important theoretical value and practical significance for improving the economic benefits of crops.In order to solve the problems of large workload and poor classification effect in the existing research,this study combines the active learning strategy of manual intervention labeling,and proposes a semi-supervised active learning method for wheat leaf disease and pest region segmentation.By extracting boundary features,color space features and texture features,the feature vector is composed,and the transductive support vector machine is used as the classifier.In each round of training,the k-means algorithm is used to randomly generate the samples to be labeled,and the human-computer interaction system is used to intervene the sample labels to improve the classification efficiency.This paper constructs a LM neural network model for classifying pests and diseases based on a new disease region segmentation method.In the image segmentation experiment,the proposed model is compared with the classical small sample SVM model and the semi-supervised TSVM model.The segmentation effect ranks first,which proves the superiority of the proposed segmentation strategy.In the experiment of pest identification,the model verification results showed that the recognition accuracy of LM neural network classification model using semi-supervised active learning method was 93.75%,which has good classification effect and can provide effective basis for subsequent pest control.

关 键 词：半监督学习 主动学习 TSVM 病虫害叶片分割 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]