基于图像处理和支持向量机的苹果树叶部病害的分类研究

发布时间：2018-01-30 21:41

本文关键词： 图像采集灰度直方图检测与分割特征向量识别准确率　出处：《西安科技大学》2017年硕士论文　论文类型：学位论文

【摘要】：近年来,随着我国经济的不断发展,农业也随之迅速发展起来,人们对农作物的产量也越来越重视。目前,如何提高农作物的产量是人们主要讨论的一个问题。病害图像的检测与识别作为病害图像系统的组成部分,对农作物病害的诊断与防治起着至关重要的作用,同时也可以提高农作物的产量,增加人们的收入。因此,对病害图像的分类识别做更进一步的研究会给人们的生活带来一定的实际意义。作物病害种类的正确识别是病害预防的前提。本论文借助图像采集装置收集了苹果树叶三类常见的病害图像。首先,对收集的病害图像进行预处理,选择直方图均衡化、中值滤波分别实现了图像的增强、去噪,再根据病害图像的灰度直方图通过双峰法确定阈值对病害图像先进行粗分割,之后再分析病害图像灰度直方图的特点对其进行精分割,并对图像的分割算法做相应的改进,从而完成病害图像的检测与分割。其次,通过分析对比几种彩色空间优劣,选择RGB和HSI彩色空间中提取分割出的病害图像的颜色特征,再依据灰度共生矩阵提取分割出的病害图像的纹理特征,并借助主成分分析法选取最具代表性的特征值对问题进行分析研究。利用这些计算简便且能反映病害本质特点的特征来组成特征向量,并构建样本特征数据库。最后,通过两种分类方法的对比,选择更适合本文病害图像分类的支持向量机模型,再通过设计分类算法,选择粒子群优化算法对支持向量机模型的参数进行优选,且对比不同参数下病害图像的识别准确率,选取识别准确率较高的参数建立病害图像分类识别模型。借助SPSS软件和MATLAB编程进行实验。利用Fisher判别分析法对病害图像进行分类,分类识别率为92.667%.再运用LIBSVB软件包和支持向量机模型对病害图像进行分类。首先,对本文中提取的病害图像的29个特征值不进行优化得到,病害图像的分类识别率为89.3939%.其次,借助主分量分析的方法选择了9个具有代表性的成分进行实验,分类准确率为92.4246%.最后,对模型的参数错分惩罚常数c和非负的松弛项g进行优选,得出当c(28)1618.28,g(28)039866.0时,此模型对病害图像的分类准确率为96.969%.达到了预期的结果。
[Abstract]:In recent years, with the continuous development of our economy, agriculture has also developed rapidly, and people pay more and more attention to the production of crops. How to improve the yield of crops is a major issue discussed by people. As a part of disease image system, the detection and recognition of disease image plays an important role in the diagnosis and control of crop diseases. At the same time, it can also increase the yield of crops and increase people's income. Further research on classification and recognition of disease images will bring some practical significance to people's lives. The correct identification of crop diseases is the premise of disease prevention. In this paper, we collect apple with the help of image acquisition device. Three kinds of common disease images of fruit tree leaves. First of all. The disease images were preprocessed, histogram equalization and median filter were used to realize image enhancement and denoising respectively. Then according to the disease image gray histogram through the double peak method to determine the threshold value of the disease image first rough segmentation, and then analyze the disease image gray histogram characteristics of the fine segmentation. And the image segmentation algorithm is improved to complete the disease image detection and segmentation. Secondly, through the analysis and comparison of several color space advantages and disadvantages. Select the color feature of the disease image extracted from RGB and HSI color space, then extract the texture feature of the disease image according to the gray level co-occurrence matrix. With the help of principal component analysis (PCA), the most representative eigenvalues are selected to analyze and study the problem. The characteristics which are simple and can reflect the essential characteristics of the disease are used to form the eigenvector. Finally, through the comparison of the two classification methods, the support vector machine model which is more suitable for the classification of disease image is selected, and then the classification algorithm is designed. Particle swarm optimization algorithm is selected to optimize the parameters of support vector machine (SVM) model, and the recognition accuracy of disease images under different parameters is compared. The classification and recognition model of disease image is established by selecting the parameters with high recognition accuracy. The experiment is carried out by means of SPSS software and MATLAB programming. The disease image is classified by Fisher discriminant analysis. The classification recognition rate is 92.667. then using LIBSVB software package and support vector machine model to classify the disease image. First. The 29 eigenvalues of the disease image extracted in this paper are not optimized. The classification and recognition rate of the disease image is 89.39. Secondly. With the method of principal component analysis, 9 representative components were selected for experiment, and the classification accuracy was 92.4246. Finally. The model parameters are optimized for the penalty constant c and the non-negative relaxation term g, and the results show that when cn281618.28g / g) 039866.0, the model parameters are deviated from the penalty constant (c) and the non-negative relaxation term (g). The classification accuracy of the model is 96.9699.The expected result is achieved.
【学位授予单位】：西安科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：S436.611;TP391.41

【参考文献】