【作者】 胡秋霞；

【导师】 何东健；

【作者基本信息】 西北农林科技大学 ， 农业电气化及自动化， 2013， 博士

【摘要】 基于图像分析的植物叶部病害识别技术研究，对有效防治农作物病害的发生，提高农作物的产量、减少农药对农产品和环境的污染，均具有重要的现实意义。本研究在国内外研究成果的基础上，以提高农作物病害识别准确率和效率为目标，以苹果、黄瓜和辣椒等作物病害为例对植物叶部病害的图像分割、特征提取、病害诊断识别方法等关键技术展开系统研究。论文的主要工作和创新点如下：（1）针对植物病斑彩色图像噪声大和边缘模糊等特点，提出了一种基于水平集和加权颜色信息的改进C-V模型。借助水平集方法对病斑图像的RGB分量图像颜色信息取加权值，以差分图像能量作为能量函数最终值，以适应不同病害种类。试验结果表明，经过RGB加权的辣椒缺水病斑图像用3R-B图像模型，苹果锈病病斑图像用3R-G图像模型自动分割，比传统C-V模型分割效果好，性能明显提高，抗噪性和可扩展性好。（2）针对加权颜色信息C-V模型运行时间长及人工选取权重参数难等问题，提出了基于相似度改进C-V模型的植物叶部病斑彩色图像分割方法。通过计算病斑彩色图像R、G、B通道的前景与背景像素均值的比例，作为病斑图像R、G、B通道的能量权值；每次迭代中将符号距离函数中小于外部能量区域符号距离平均值的距离值归零，得收敛分割曲线。试验结果表明，该方法平均准确率分别高于基于加权颜色信息的C-V模型和传统C-V模型0.42和43.55个百分点，且抗噪性好；对于高分辨率彩色图片，平均运行时间不到这二者的1/1000，大大提高了算法效率。（3）针对颜色矩运行时间长及分类率不高的问题，提出了基于颜色矩和小波变换的加权特征提取方法，该方法分别对H、S和V通道图像进行小波分解根据其1-3阶矩特征得到每个子图的能量系数，将其作为特征的权重。试验显示，用支持向量机多项式核函数，小波基为bior2.4时提取特征的识别正确率最高，花叶病、锈病和斑点落叶病平均识别正确率为89.78%，与基于颜色矩和小波的特征提取方法的识别正确率（25.67%和80.14%）相比，显著提高了识别正确率。（4）针对小波分解进行特征提取通常仅利用了图像低频子带信息的缺点，提出了基于YUV和小波包的多通道特征提取算法，以全面分析图像纹理特征。该方法对输入图像的Y、U、V通道子图像进行小波包分解，以Y、U、V分量子图像各自小波包能量特征和分量子图像之间的小波包能量特征作为病害图像特征向量。试验结果表明，当小波基为haar，支持向量机为多项式核函数时，花叶病、锈病和斑点落叶病的平均识别正确率为89.10%，比基于小波的特征提取方法识别正确率提高12.16个百分点。（5）针对SVM模型中参数难以确定的问题，提出了基于遗传算法支持向量机模型（GA-SVM）的苹果叶部病害识别方法。该方法用遗传算法自动获取SVM参数，GA参数用旋转正交方法得到，获得合适的GA-SVM模型后，将基于颜色矩和小波变换的加权特征作为特征向量进行分类。试验结果表明：当粒子数M=50，交叉率Pc=0.7，变异率Pm=0.05，迭代次数G m=100时，花叶病、锈病和斑点落叶病的平均识别正确率为92.39%，比SVM模型提高了5.21%识别正确率。（6）针对SVM模型中参数难确定及遗传算法复杂问题，提出了基于粒子群算法支持向量机模型（PSO-SVM）的苹果叶部病害识别方法。该方法用粒子群算法确定SVM模型的惩罚因子和核参数的最优数值，得到合适的PSO-SVM模型后，以基于颜色矩和小波的加权特征作为特征向量进行分类。经旋转正交试验探明较优粒子群算法参数为c1=c2=2.05，粒子数m=20时，病害识别正确率较高，苹果花叶病、锈病和斑点落叶病的识别正确率分别为90.24%、87.26%和85.23%，平均运行时间比GA-SVM方法减少10.86个百分点，提高了识别效率。更多还原

【Abstract】 The image analysis-based research on plant leaf disease diagnosis recognitiontechnology has a great significance to prevente the occurrence of crop disasters effectively,reduce the impact of the crops diseased, increase crop yield and reduce pesticides onagricultural products and environmental pollution. On the basis of adequate summary aboutthe advanced achievements at home and aboard, in this paper systematical research has beenmade on the key technology of image segmentation, feature extraction and disease diagnosisrecognition, which relyed firmly on the goal of improving recognition accuracy and efficiencyof crop diseases and its cases study of apple, cucumber and capsicum disease. The chief workand major innovations are as follows:(1) In view of the characteristics of noise and blurred edges for plant lesion colorimages, a novel C-V model named WCCV based on level set and weighted color informationwas proposed in this paper and applied to plant lesion image segmentation.The WCCVsegmentation model is suited to different disease identification and can identify lesion diseaseautomatically. Experimental results show that the proposed model has better property than thetraditional C-V model, and have many advantages such as anti-noise and scalability propertieson3R-B image model for capsicum water shortages disease and3R-G image model for applerust disease.(2) According to the flaw of WCCV model, an improved C-V model was proposed inthis paper and applied to plant leaf lesion image segmentation. At first, a point in lesion areais selected, and the mean value calculated from its3×3territory is used to compute thesimilarity between this point and other points in the image, and the foreground andbackground are determined. Then the ration of foreground and background mean pixels fromR,G, B channels is used as weighted value of these three channels, respectively. Finally, thelevel set function is solved iteratively to gain the segmentation contour, while let the distancevalue less than the average distance value of external energy region be zero in the signeddistance function.Experimental results show that the average segmentation accuracy of theproposed method is0.42%and43.55%higher than that of WCCV model and the traditionalC-V model. For large pixel pictures, the average running time of the proposed method is lessthan1/1000of WCCV model and the traditional C-V model. It reduces the running time andimproves the efficiency of algorithm execution greatly.(3) Aiming at the shortcoming of the long running time and not high classification rate of color moments, the weighted feature extraction method based on color moments andwavelet decomposition was proposed. At first, the original image is transformed into H, S andV-channel images to get the wavelet decomposition of each channel subgraphs. Secondly, thedecomposition subgraphs of the integration color and wavelet features are gained and thefirst-order, second-order and third-order moments features of the decomposition subgraphsare calculated. And the energy coefficients are as the weighted features value of the colormoments of the subgraphs and finally the feature vector is gained. The experimental resultsdemonstrate that when the polynomial kernel function-based SVM(support vector machine)and the bior2.4-based wavelet are used, the proposed method recognition correct ratios are88.57%,88.46%, and92.31%for apple mosaic virus, apple rust and apple alternaria leaf spot,respectively, and the average recognition ratio is89.78%. The proposed approachsignificantly improves the recognition correct ratio compared to the25.67%and80.14%recognition correct ratios for color moments-based and wavelet-based feature extractionmethods, respectively.(4) A YUV and wavelet packet-based multi-channel feature extraction algorithm wasproposed. Due to the drawback of the wavelet decomposition-used the feature extractionmethod is normally only using the low-frequency sub-band information of the image, so theuse of wavelet packet decomposition for all frequency channels is a comprehensive analysisof images texture features. Firstly, the input image is converted into the Y, U, V channelssubgraph, respectively. Then the wavelet packet energy features of the Y, U andV-component images are calculated. Finally the wavelet packet energy features among thesub-quantum images are obtained as the disease image feature vector. The experimentalresults demonstrate that when using the polynomial kernel function-based SVM and the haar-based wavelet packet, the average recognition correct ratio is89.10%.Compared with thefeature extraction method based on wavelet transforming, the proposed approach boosts12.16percent recognition correct ratio.(5) The genetic algorithm (GA) is used to select the parameters of the SVM methodautomatically and the orthogonal method is utilized to determine the best GA parameters.Firstly, the color moments and wavelet-based features of apple disease leaf images areextracted as feature vectors. Then the proposed GA-SVM method is used to classify appleleaf disease images.The orthogonal experimental results demonstrate that the proposedGA-SVM model recognition correct ratios are94.47%,91.44%, and91.26%for apple mosaicvirus, apple rust and apple alternaria leaf spot, respectively. And the average recognitioncorrect ratio is92.39%.Compared with the recognition method based on SVM, the proposedapproach shows5.21%higher recognition correct ratio than that of SVM.(6) Considering the difficulty of parameter determination in the original SVM and thecomplexity of genetic algorithm, the particle swarm optimization algorithm (PSO) is used toselect the parameters of the SVM automatically and obtain an optimized function. The PSO algorithm has no crossover and mutation operation, less parameters and easier to use.And itsprinciple is more simple than that of GA. Firstly, the color moments and wavelet-basedfeatures of apple disease leaf images are extracted as feature vectors. Then the proposedPSO-SVM model is used to classify apple leaf disease images. The orthogonal experimentalresults show that the recognition correct ratios are90.24%,87.26%, and85.23%for applemosaic virus, apple rust and apple alternaria leaf spot, respectively. Compared with therecognition method of GA-based, the proposed approach costs10.86%less processing timethan that of GA-SVM.更多还原