【作者】 宋凯；

【导师】 纪建伟；

【作者基本信息】 沈阳农业大学 ， 农业电气化与自动化， 2008， 博士

【摘要】 为保证有效合理地施用农药防治农作物病害,农业生产者必须准确的获取作物的生长信息,这样,农业生产者可根据获得的病害信息快速、准确的诊断受害作物的病因以及受害程度,因病治宜。随着计算机处理能力的不断增强,以及图像处理与识别技术的快速发展,数字图像处理与识别技术在农业中的应用越来越广泛,并将成为实现农业信息化与自动化的重要技术力量。农业信息采集工作量巨大,信息的实时性和准确性是农业生产和科学研究领域普遍关注的问题,如何及时快速地进行农作物病害的准确判断一直是计算机技术面向农业领域研究工作的一项重要内容。为此,本文以计算机视觉技术为重要技术手段,综合运用图像处理和植物病理学方面的知识,研究玉米和黄瓜病害的识别和诊断。首先,根据病害叶片的采样要求,利用光照系统和计算机图像处理装置进行病害样本的图像采集。但是无论采用何种装置,采集到的图像往往不能令人满意。针对所采集的图像包含噪声的问题,讨论图像去除噪声的方法。在去除噪声方面论述了常用的几种消噪方法,很多消噪的方法可以很好的去除噪声对图像的影响,但是在消噪的同时也弱化了图像中目标区域的边缘,不利于基于边缘的图像分割算法的使用。因此本文采用Winer滤波来对病害图像进行去除噪声,同时采用多尺度Retinex彩色复原图像增强算法对图像进行增强处理,改善图像质量。经上述处理后,图像质量和显示效果得到明显改善,符合实验要求。其次,深入研究了图像分割的各种方法,仔细研究了病害图像的特点,将聚类分析引入到图像分割中,分析比较了硬C—均值聚类和模糊C—均值聚类分割算法的特性,采用快速模糊C—均值聚类算法,对病害图像进行分割,并通过实验验证了这种算法在聚类优化性能不变的前提下,可以使运算的开销降低,从而使得分割耗时明显地减少。本文根据Kingsbury提出的具有近似位移不变性和良好方向选择性的Q-shift DTCWT变换理论设计了基于统计性特征和系数特征的提取算法。提取了病斑图像的周长、面积和形状参数等特征,然后对所获得的特征值进行标准化,并进行病害图像的分类判断,以获得病害识别的精确性。对训练样本特征提取阶段的结果进行训练SVM分类器,并应用训练好的SVM分类器进行分类识别检测。在病害图像预处理和特征提取阶段采用了不同的方法并在不同的Video中与SVM分类法进行了大量的组合测试。结果表明,本文提出的病害识别算法不仅具有较好的鲁棒性,而且能有效提高识别率和降低误识别率。采用3层完全结合方式的Bp神经网络来建立农作物病害的诊断模型,同时将模拟退火算法和粗粒度并行遗传算法结合起来,既综合了遗传算法和模拟退火算法的优点,又加快了一般模拟退火遗传算法的搜索速度,对所建立的BP神经网络进行优化。优化完成后,网络的诊断能力和运算速度得到增强。采用VC++程序设计软件编写程序,形成了基于计算机视觉的农作物病害识别系统。本文从算法理论研究入手,以计算机图像处理为技术手段,以VC++语言为编程语言,综合运用计算机视觉技术、人工神经网络、小波变换、支持向量机和统计模式识别方法,对作物病害图像的处理和诊断技术进行了研究。更多还原

【Abstract】 In order to ensure the effective application of pesticides over crop diseases control, agricultural producers must achieve the crop growth information accurately. Based on the acquired information, they can take rapid diagnoses on the causes and extent even the measures to control the diseases. With the rapid progress of computer processing and image recognition, much more fields of agriculture are using the technology to realize digital processing and automation. However, the data collection of crop growth is a hard job and the accuracy and real-time of information is always a concern in the field of agricultural production and scientific research. Therefore, it becomes so important to judge the type of crop diseases accurately by using computer technology to guide the agricultural production. Hence, this paper makes a study of corn and cucumber diseases recognition and diagnoses by different image processing and pathology based on computer vision technology.First, use illumination system and image procession device to collect diseases samples according to the requirements of infected leaves. But the devices used to be unsatisfactory in the collection. The noise always affects the quality of the image. Several common methods are used to remove the noise, but they also weaken the image on the brink, which is not useful in the image segmentation algorithms. In this paper, Winer filter and multi-scale recovery Retinex color image enhancement algorithms are used to improve the picture quality. And the facts prove that the image quality and effects are much more improved after the treatment.Second, study profoundly over image segmentation methods and the characteristics of all kinds of disease images. Cluster analyses are introduced in image segmentation to analyses and compare C-means clustering and Fuzzy C-means clustering segmentation algorithm characteristics. Experiments show that this method can reduce the computer cost and make segmentation time consuming.This article proposes extraction algorithm according to Kingsbury’s idea of approximate displacement invariability and directionally selective Q-shift DT CWT transform theory based on his statistical and coefficient characteristics. The data of perimeter、area and shape over the infected leaves is collected to standardize it for further classification to ensure the accuracy. The trained samples withdrawn are put into the SVM trainer to have further feature recognition examination. Different methods are used and massive combined tests are carried on in different Video and SVM classification in the pretreatment and feature extraction stage. It shows that the method proposed in this article is not only effective in robustness but also significant in promoting recognition rate.To adopt 3-layer Bp neutral network to establish a crop disease diagnostic model and combine the annealing algorithm with the coarse grain parallel genetic algorithm can both keep the merit of synthesized genetic algorithm and speed up the searching time in genetic algorithm. After the optimization, the network diagnosis ability and the computing speed will be improved efficiently.To adopt VC+ programming software device can set up a system based on computer vision over crop disease recognition.This paper finishes a research over crop diseases recognition by using algorithm technology based on computer image processing, VC+ language, comprehensive computer vision device, artificial neutral network, wavelet transformation, support vector machines and statistical model recognition methods.更多还原