【作者】 王娜；

【导师】 李少昆； 王克如；

【作者基本信息】 石河子大学 ， 农业， 2009， 硕士

【摘要】 本文研究利用计算机图像处理算法和模式识别技术实现作物病害自动识别诊断,以解决目前植保技术人员分配不均,已开发病害专家系统无法满足用户需求等问题,为进一步开发具有智能化的植物病害诊断专家系统提供了必要的前期研究。以常见的六种玉米叶部病害：玉米灰斑病、弯胞菌叶斑病、小斑病、普通锈病、大斑病、褐斑病为研究对象,大田环境条件下采集玉米叶部病害图像样本,根据病害图像特点,对病害图像样本进行预处理得到目标图像,提取目标图像信息特征,选择用于病害诊断的特征向量,利用模式识别方法对病害图像进行自动识别诊断。本文从图像采集、图像预处理、特征提取、模式识别等几个方面都进行病害图像识别研究,取得了较理想的研究结果：1.制定了叶部病害采集规范。本文中研究了不同品牌数码相机、不同分辨率、不同物距、参照物引入与否等成像环境因素对图像采集和后期图像处理的影响。在此基础上规范了叶部病害图像采集环境和操作步骤,对保障叶部病害图像识别研究的质量控制和研究成果的共享具有重要的理论意义和实用价值。2.根据玉米叶部病害图像特点,确定了处理效果好、普适应强的叶部病害图像流程。研究比较了利用邻域平均和中值滤波算法去除图像噪声,利用图像锐化技术和对比度增强算法增强图像,并综合应用基于H阈值分割、图像灰度变换、迭代二值化、图像形态学运算、轮廓提取等算法处理病害图像,抽取病斑。3.综合提取玉米叶部病害图像纹理、颜色、形状等多向量特征参数。利用灰度共生矩阵算法和计盒维数法提取病害图像的墒、能量、惯性、对比度、相关性、相关信息测度、分形维数等纹理特征值；综合应用RGB颜色空间、HIS颜色空间、HSV颜色空间、YCbCr颜色空间提取病斑图像的RGB颜色空间/HlS颜色空间/YCbCr颜色空间的各颜色分量、各阶颜色矩等颜色特征值；应用轮廓提取和区域跟踪算法计算病斑病斑周长、面积、圆度、形状因子、离散指数、等效圆半径、内切圆半径等形状特征值。4.为提高识别效率和精度,采用遗传算法自众多原始特征中优选出相关信息测度、分形维数、归一化蓝色颜色分量b值、cb、颜色矩、病斑周长、病斑面积、圆度、形状因子等9个独立、稳定性好、分类能力强的特征向量用于病害识别。5.利用费歇尔判别分析方法和贝叶斯判别分析方法分别识别玉米叶部病害,病害识别结果显示,应用费歇尔判别分析方法判别六种病害的正确识别率均在85%以上,贝叶斯判别分析方法对六种病害的止确识别准确率均在90%以上。研究为进一步开发具有智能性的作物病害诊断专家系统提供了基础,促进图像识别技术在农业工程领域的应用具有一定的意义。更多还原

【Abstract】 The target of the research is to diagnize main maize leaf diseases by machine vision. To improve the actuality of uneven distribution of plant solution technical personnel, and disease expert system developed can’t meet customer requirements. And it is necessary for further development of intelligent disease expert system as a basic reseach.The research took gray spot, curvular leaf spot, bipolaris maydis, rust, northen blight, brown spot as examples. The diseases pictures of different varieties and periods were taken in fields. According to the features of crop maize disease, image processing algorithms were applied to obtain object image. Then image features were extracted, and approximate features were chosen. Finally, pattern recognition technology was used to identify diseases automatic.The paper studied adopting image processing, characteristic extraction and pattern recognition to identify disease, and the desired results were achieved.1. The standards of disease image acquisition were established. Different imaging factors such as camera brand, focus distance, imge resolution and reference object were stuied to address the effects on the recognition of the disease images. Disease image collection method was regulated. And laid foundations for the standardization disease image collection framework.2. Optimum and universality method of image processing was determined according to features of maize disease image. Agorithms of neighborhood average and middle filter were used to eliminate noise; and algorithms of image sharpening and image contrast were used to enhance image; then methods of threshold segmentation based on hue, iteration binarization, image morphological operation and contour extraction were adopted for image segmentation.3. Multivariable characteristics such as texture, color and figure characteristics were extracted. Applying gray level co-cccurence matrix and box counting methods to computer entropy, energy, inertia, contrast, relativity, informationization measure, fractal dimension et al; color components based RGB, HIS, YCbCr, color moments were computed by utilization of color models of HSV; the algorithms of contour extraction and edge detection were applied to extract shape features, suah as girth, area, rotundity, figture factor, scatter index, equivalent circle radius, inscribed circle radius.4. In order to improve recognition efficiency correctness, Genetic algorithm is used to choose approximate features, and 9 optimum and effective image features including informationization measure, fractal dimension, color components b and Cb, color moment, girth, area, rotundity, figture factor were selected as recognition features from 29 primordial features.5. Pattern recognition methods of fisher discrimination analysis and bayes discrimination analysis were applied to recognize diseases. The results indicated that the recogniton precision of six kinds of maize disease recognition is higher than 85% by fisher discrimination analysis, and higher than 90% by bayes discrimination analysis.The paper provided essential theoretical evidence and technological basis for the futurer development of plant disease recognition system. And it improved computer image processing technology in agriculture engineering field.更多还原