【作者】 王访；

【导师】 廖桂平；

【作者基本信息】 湖南农业大学 ， 作物信息科学， 2013， 博士

【摘要】 叶片是作物的重要生理器官之一,其图像能有效反映作物营养素缺失和病害种类及影响程度。在数字农业中,基于计算机和数学方法的叶片图像处理是研究上述问题的重要途径。如何通过数学算法和机器智能提取叶片图像的有效信息,成为了对作物营养素缺失、病害影响的叶片诊断的关键问题。目前的研究大多集中于作物叶片的颜色、形状特征,对叶片纹理信息的研究较少。而纹理是叶片固有的本质特性,不易受环境因素影响,只有当作物受营养缺失、病害影响时,其叶片灰度图像的纹理会有相应的改变,因此,作物叶片的纹理特征也是研究上述问题的理想对象。多重分形理论是一种描述图像纹理特征的重要手段,近年来在图像处理领域得到了很好的应用。本文利用多重分形理论,针对叶片图像不平稳的特点,提出了一种叶片灰度图像多重分形特征描述方法,得到了一些具有鲁棒性的纹理描述因子,并利用这些特征对上述问题展开了研究。旨在通过机器智能为作物叶片缺素和病害进行无损诊断奠定理论基础。1.提出了图像平稳性的定义,给出了两种检测图像平稳性的方法。对图像平稳性的讨论是决定用何种多重分形方法提取图像特征的前提工作。由于标准多重分形算法都是基于图像平稳测度而提出的,对于非平稳测度将会得到不准确的结果。而受缺素和病害影响的叶片图像不同于其他图像,其局部灰度值易发生突变,可能产生不平稳测度。针对这一现象,定义了二维灰度图像的平稳性,提出了两种检测平稳性的方法,并利用一组由分形高斯噪声生成的图像和一组由分形布朗运动生成的图像验证了定义和检测方法的有效性。6种玉米病害叶片图像的平稳性检验结果表明它们都是非平稳的。2.提出了基于多重分形去趋势波动分析(MF-DFA)的纹理图像特征描述方法。采用合适的多重分形特征提取方法是获取叶片图像纹理特征有效信息的保障。针对叶片图像非平稳性和标准多重分形不能解决非平稳测度的缺陷,基于多重分形去趋势波动分析(MF-DFA),提出了一些描述纹理图像特征的算法,获得了全局图像灰度值序列的广义Hurst指数h(q),二维图像表面全局广义Hurst指数H(q)和图像局部广义Hurst指数LHq等新的纹理描述符。一方面,与单一分形维数、标准多重分形谱、基于灰度共生矩阵法的角二阶矩、对比度、熵和自相关系数等传统的纹理特征因子进行对比实验,结果表明提出的h(q)和H(q)具有最好的抗噪性(平均误差<2%)、抗压缩性(平均误差<5%)、抗模糊性(平均误差<7%)。另一方面,与传统的基于差分盒子法的局部广义分形维数及基于三种容量测度的Holder指数进行分割对比实验,结果表明在相同的分类器下,提出的LHq具有最好的分割效果(对不同纹理图像的平均识别率>90%),且具有最好的抗噪性(平均误差率<10%)。3.提出了基于局部多重分形去趋势波动分析(LMF-DFA)的图像分割方法。图像分割问题是正确反映叶片图像奇异区域,并对其进行缺素、病害诊断的关键问题。针对缺素、病害的叶片图像局部灰度不平稳特征,提出了一种基于LMF-DFA的图像分割方法。该方法以LHq为纹理描述符,以LHq构成子图的计盒分形维数f(LHq)为依据对指定区域进行分割。缺镁素、钾素的油菜叶片图像和玉米小斑病、灰斑病、弯孢菌叶斑病、圆斑病、锈病和褐斑病叶片图像的分割实验表明该算法能有效地分割叶片图像中的奇异区域。通过与现有流行的基于容量测度的多重分形谱分割法及经典的模糊C均值聚类法进行对比实验,结果表明所提方法分割效果最好,既能正确反映上述叶片缺素和病害区域的位置,也能准确识别非缺素、病害的区域,同时具有良好的抗噪性。4.建立了基于叶片图像的多重分形特征的油菜氮营养诊断模型。基于叶片特征的模型构建是研究作物营养诊断的核心问题。一方面从定量的角度为不同施氮水平下的油菜氮含量建立了回归模型,利用叶片图像的多重分形特征参数为自变量预测氮含量的相对均方根误差最低为10%-20%。另一方面,对大田环境下采集的不同施氮水平下的油菜叶片进行了定性诊断。对基部、中部和顶部及三个部位混合样本进行分类识别。结果表明基于基部叶片和中部叶片的识别效果明显优于顶部叶片。对于混合叶片样本,在5折交叉检验方法下,以支持向量机核方法和随机森林为分类器的平均诊断准确率分别为94.03%和94.90%。更多还原

【Abstract】 Leaf, as a vital organ of crops, can reflect growth conditions of the crops directly. The type of nutritional deficiency and plant disease can be mirrored by the images of the leaves. The images can also provide critical clues for identifying and diagnosing different level of nutrition. In the research of digital agriculture, it is an important way to solve above problems to process the leaf digital images based on mathematics and computers. How to get the useful information from the leaf images have become the key issues to locate, recognize and diagnose the singular region in the images impacted by the nutritional deficiency and diseases. Current studies are focused on the leaf color and shape characteristics, few studies on the texture information. However, as an inherent nature and characteristics of leaf, the leaf’s texture remains relatively stable in the crop growth and less susceptible to outside influence. Meanwhile, the leaf’s texture will change when the leaves are affected by nutrient deficiency and disease. By this token, the texture features of leaf image are ideal object for researching the above problems. Multifractal theory is an important tool to describe the texture features, which is widely applied in the field of image processing. Targeting at non-stationary trait of the leaf images, some robust texture descriptors are proposed in our paper by the multifractal theory to study the above problems. It should lay foundation for diagnosing the corps with nutritional deficiency and diseases nondestructively by machine intelligence.A definition of image stationary and two methods of detecting the image stationary are proposed. The research of the images stationary problem is the premise concern in determining which multifractal technologies to extract the texture feature for the leaf images. Since the existed multifractal methods were proposed based on stationary measures, they cannot solve the non-stationary problems. Unlike the other images, the leaf images with nutritional deficiency and diseases easily lead to mutation of local gray values and may produce not smooth measures. In allusion to this phenomenon, we define the image stationary tentatively and present two methods of stationary detection based on the stationary of one-dimension series. The investigation by two groups of synthetic images generated by fractional Gaussian noise and fractaional Brown motion surface, respectively, shows that the definition and the detection methods of image stationary are effective. In the experiment of detecting the stationary for six corn disease leaf images by the above methods, it shows that they are all non-stationary.Some texture characterization algorithms based on multifractal detrended fluctuation analysis (MF-DFA) is proposed. Employing right feature extraction method by the multifractal theory is the important safeguard for getting the useful texture information of the leaf images. Targeting at non-stationary trait of the leaf images and the problem of the non-stationary measures cannot be solved by standard multifractal analysis; we propose a texture characterization method based on multifractal detrended fluctuation analysis (MF-DFA). Three new texture descriptors, namely, the generalized Hurst exponent of gray series for global image, denoted as h(q), the generalized Hurst exponent of two-dimension surface for global image, denoted as H(q), and the generalized Hurst exponent of local two-dimension surface for each pixel in the image, denoted as LHq, are obtained. Next, we test the three exponents by two groups of experiments. On one hand, by comparing with the traditional texture descriptors, such as mono-fractal dimension and multifractal spectrum based on mono (multi)-fractal, four statistics calculated by gray occurrence matrix method, the results demonstrate that the proposed h(q) and H(q) have the best noise immunity (the average error is less than2%), best compression resistance (the average error is less than5%) and best anti-ambiguity (the average error is less than7%). The three kinds of errors are less than responding errors calculated by the other texture descriptors significantly. On the other hand, both the proposed texture descriptor LHq and other two multifractal indicators, which are Holder coefficients based on capacity measure and generalized multifractal dimension Dq based on multifractal differential box-counting (MDBC) method have been compared in our segmentation experiments. The first comparative experiment indicates that the segmentation results obtained by the proposed local multifractal detrended fluctuation exponent are as well as the MDBC-based Dq and superior to the Holder coefficients significantly. The results in the second comparative experiment of noise immunity demonstrate that the proposed method can distinguish the texture images more effectively and proazvide more robust segmentations than the MDBC-based Dq significantly.A novel image segmentation algorithm based on local MF-DFA (LMF-DFA) is proposed. The image segmentation problem is the key problem for illustrating the singular regions in the leaf images. In allusion to non-stationary trait of the leaf images with nutritional deficiency and diseases, we propose a novel image segmentation algorithm based on LMF-DFA. In the proposed algorithm, the local generalized Hurst exponent LHq is calculated firstly. And then, box-counting dimension f(LHq) is calculated for the sub-images constituted by the LHq of some pixels, which come from a specific region. Consequently, series of f(LHq) of the different regions can be obtained. Finally, the singular regions are segmented according to the corresponding f(LHq). The method is used to locate the lesion regions for the six corn disease leaf images, namely, Southern corn leaf blight, Gray leaf spot, Curvularia leaf spot, Round leaf spot, Rust disease and Brown spot. Meanwhile, both the proposed method and other two segmentation methods--multifractal spectrum based and fuzzy C-means clustering have been compared for the six disease images. The results indicate that the proposed method can recognize the lesion regions more effectively and provide more robust segmentations.A kind of diagnostic model for rapeseed nitrogen based on the multifractal feature of leaf image is proposed. Model construction is the kernel problem in researchingthe crop nutrition diagnosis based on leaf characteristic. On the one hand, we construct some regress models for the rapeseed nitrogen content under the different levels. The least relative square mean root error of predicting the nirtrogen is about10%-20%by using the multifractal parameters as the independent variables. On another hand, a qualitative diagnostic model is proposed for identifying different nitrogen level of rapeseed leaves. In the model, using the selected multifractal parameters as image characteristics, we take diagnosis analysis for base, middle and top of rapeseed plant with different nitrogen levels, which are collected from the field environment. The experiment result shows that the best diagnose accuracy comes from the base of the rapeseed leaves. It is explained that the base leaf is most sensitive to the nitrogen deficiency. In the diagnose of the mixed samples, the accuracy reaches94.03%和94.90%under support vector machines and kernel methods and random forests methods by the5-fold cross validation.更多还原