【作者】 任东；

【导师】 于海业；

【作者基本信息】 吉林大学 ， 农业机械化工程， 2007， 博士

【摘要】 支持向量机(support vector machine，简称SVM)是一种新的模式识别方法，它采用结构风险最小化(SRM)原理，兼顾训练误差和泛化能力，在解决小样本、高维非线性、局部极小值等模式识别问题中表现出特有的优势。本文以温室植物病害为主要研究对象，以支持向量机技术为重要技术手段，并将地面环境信息引入，进行温室植物病害的诊断研究。本文提出了一种新的核函数——线形组合高斯核函数核，并将该核函数应用于支持向量机方法中，并使用该方法对北京地区黄瓜病害图像进行了识别分类，达到更好的分类效果。在病害图像特征提取中，采用基于启发式搜索策略的支持向量机的Wrapper方法—回归特征消去方法(SVM-RFE)对原始特征进行选择。在分类器设计方面，提出采用模糊多类别分类器，该方法很好的解决了“一对一”和“一对多”分类器中出现的不可识别区域问题。在设计病害识别系统方面，将植物生长环境信息对植物病害发生的影响引入病害诊断中，并将这种影响量化为影响因子，作为支持向量机分类器的加权系数，来修正最后的识别分类结果。本项研究为进一步开发具有使用价值的植物病害诊断系统提供了重要的理论基础和应用技术，对缩小我国在农作物自动化管理方面与发达国家之间的差距，促进现代技术在我国农业工程领域的应用具有重要意义。更多还原

【Abstract】 It is a effective diagnosis method for plant disease detection that using computer image processing and pattern recognition technology to identify plant diseases. During actual disease diagnosis, for a particular disease, here is not usually a lot of disease samples. many pattern recognition methods (including artificial neural networks) can not be good for plants identification, which affect the accuracy of the diseases diagnosis. SVM (support vector machine) is a new method for pattern recognition, which is the structural risk minimization (SRM) principle, balance training error and generalization ability, have solved small sample, nonlinear, High Dimension local minimum value of pattern recognition problem, which demonstrated the unique advantage.In theory, SVM is a stable classifier, has good generalization ability. However, In the specific application process, which need to overcome some difficulties: choicing the kernel function and classification parameters; promoting two-types problems into a wide range problems. which impacts stability and generalization ability of SVM method. This paper is based on the foregoing, regarding greenhouse plant diseases as the mam study. using SVM technology for the important technical means, and introducing environment information to explore specific issues of potential, capabilities and prospects for support vector machines on greenhouse plant disease diagnosis. Main contents is including:(1) reasonably choice and design kernel function if an important part of Support Vector Machine, the different kernels function represent different nonlinear mapping that Using Support Vector Machine to solve nonlinear Classification problems. So kernel function Support Vector Machine which is easy to achieve nonlinear algorithm.This paper presents a new kernel function------linear combination Gaussian kernelfunction support vector machines. Firstly, have studyed Gaussian kernel algorithm, to improve the Gaussian kernel performance by Transforming the classical Gaussian kernel.Combining Gaussian kernel with several other classic nuclear linear of function into a new combination of kernel, which will improve it’s ability. the kernel is Applicated in Support Vector Machine, and use this method to recognize cucumber diseases Image in Beijing, and the results Satisfied. It is proved that the linear combination of Gaussian kernel is better than a single kernel in classification results. the artificial neural networks and other classic SVM Kernel classification results were compared, The results also show that the kernel is effective.(2) In extraction of the diseases image feature, extractiing plant disease image original characteristics with plant diseases image of shape, texture and color information. The author raised a new method that apply support vector machine theoryto recognize plant disease------Recursive Feature Elimination (SVM-RFE) to getcharacteristics of the original. The results that Experimental data was analysised show: A-SVM-RFE algorithm efficiency is the highest. H-SVM-RFE is the fastest among three SVM-RFE methods, but it lost a variables. O-SVM-RFE recognition rate is relatively low. The method is superior to other identification methods, in feature extraction, A-SVM-RFE is the best performance. Meanwhile with the traditional method, the test results show that of the new feature selection method Based on the Support Vector Machine theory is better than the traditional method(3) Plant Disease Diagnose is a multi-classification problem. it may be many diseases For a plants, it be considered to design the categorized classifier in the plant diseases identification. SVM theory is raised against two issues, so the first thing using support vector machines mathod should be completed design of multi-category classifier. There is inseparable subregion for raditional SVM multi-category classification method. Author combined the "one-on-one" method with Fuzzy Support Vector Machine to make up for the flaws. Auhtor use multi-category classification to identify greenhouse cucumber three diseases, and compared it with the "one-to-one" and "one-to-many" classifier, the results showed: the recognition rate of FSVM method is 98%. the recognition rate of "one-on-one" method is 95%. the recognition rate of "One-to-many" method is 93%, from above, FSVM method soluted the inseparable subregion issues of "one-to-one" and "one-to-many" methods, its performance is the best.(4) plant environmental information is a important factor that impact the incidence probability of plant disease, Author geted the curve of nvironmental factors affect the incidence of the disease, and the environmental information disease forecasting model by analysizing plants of environmental factor effect the probability of disease incidence. It have been introduced that plant environmental information affect plant disease incidence, and quantify the impact of factors, as weighted coefficient of support vector machine classifier, Finally amended the identification classificater,Which can greatly improve the classifier performance and eliminate diagnostic errors of image acquisition, pre-processing. The recognition rate is 96.7%, 99.6% and 97.3% after intraducing environmental information. greatly improved the classification performance, and eliminated the error that introduced in image preprocessing , to a large extent, avoid appearance of the wrong, environmental information to a large extent ensure correctness of the results of the identification. When result for SVM is closed.It is useful to further develop plant disease diagnosis system, which has provided an important theoretical basis and the application of technology. It is great significance to reduce gap between developed country and our country and to promote modern technology in the field of agricultural engineering applications.更多还原