小麦叶部病害识别方法研究及智能手机诊断系统研发

发布时间：2018-10-19 06:29

【摘要】：病虫害是影响农作物产量和品质的重要因素,如何对其进行实时监测和快速准确区分对指导农作物生产管理具有重要意义。传统的监测和区分方法是通过植保专家抽样调查、人工区分进行判断,费时费力,难以满足大面积调查的需求。随着科学技术的快速发展,植保专家利用图像处理、模式识别、遥感等技术对植物病虫害进行监测和识别,取得了显著效果。但是,已有技术和开发的传感器距离实用、低值、便携应用仍有差距。本论文以小麦叶部条锈病、白粉病为观测对象,结合图像处理和模式识别技术探索适用仪器开发的小麦病害快速识别方法,设计了一款基于Android智能手机的病害诊断系统。主要内容、创新点和结果如下：(1)研究了高通滤波、中值滤波、邻域平均法3种图像增强算法用于减少图像采集环境带来的影响；研究了3种分割算法(优化分水岭分割、自动阈值分割和水平集分割),以将病斑从小麦健康叶片中分离出来,用于提取病斑特征。从颜色、形状和纹理三个方面(共计提取23个病斑特征参数)对小麦叶部病害特征进行描述。试验结果表明：单一的图像增强算法不能达到理想增强效果,且单一的图像分割算法也不能很好地将目标区域分割出来。因此,应将图像增强和分割算法进行优化以提高其增强和分割效果。(2)研究了相关向量机(Relevance Vector Machine, RVM)、支持向量机(Support Vector Machine, SVM)口反向传播神经网络(Back Propagation Neural Network, BPNN)三种图像识别方法。本论文中选取150个不同严重程度(含轻度、中度和重度)的小麦叶部病害(条锈病、白粉病)为试验材料,以轻中度病害为重点,选取其中68个病害叶片为训练样本,提取每个病害叶片的颜色、纹理和形状共计23个特征,利用Relief算法计算病害颜色、纹理和形状中每个特征的权重(即对病害识别的贡献大小),并选取其中20个权重较大的特征作为SVM、 BPNN和RVM的输入参数,分别建立三种识别模型。通过2组试验的68个测试样本验证,结果显示：SVM、BPNN和RVM的平均识别准确率分别为86.76%、91.17%和89.71%,而对轻中度病害的识别准确率分别为86.67%、90.00%和88.33%,其中,RVM的执行效率分别是SVM和BP神经网络的7.96和31.68倍。(3)针对目前装置携带不便、价格昂贵、专业性要求高等问题,结合RVM识别算法,开发了一款基于Android智能手机的小麦叶部病害诊断系统。本文利用Sony DSC-H9相机和SAMSUNG GT-N7100手机采集白粉病和条锈病不同严重程度(轻度、中度和重度)样本各66个(白粉病和条锈病各33个),选取其中48个(白粉病与条锈病各24个)作为训练样本,其余用作测试样本；同时,改变手机采集样本像素作为另一对照组来研究像素与识别率的关系,同上安排样本分布。研究结果表明：RVM得到的平均识别率为88.89%,病害的正确识别率与采集工具有关,并与其像素成正比。因此,进行病害识别需选择像素合适的手机以得到较高的准确率。同时,经应用测试发现,识别一副病害图片可在20s内完成,能够实现对小麦不同严重程度叶部病害快速准确识别,这为植保人员田间调查提供了重要的技术支持。
[Abstract]:Pest is an important factor affecting the yield and quality of crops, and how to monitor and distinguish it in real time is of great significance to guide crop production management. The traditional method of monitoring and distinguishing is through sample sampling of plant protection experts, artificial differentiation and judgment, and it is difficult to meet the needs of large-area investigation. With the rapid development of science and technology, plant protection experts use the techniques of image processing, pattern recognition and remote sensing to monitor and identify plant diseases and insect pests. However, prior art and developed sensor distances are practical, low, and portable applications still have gaps. This paper uses wheat leaf rust and powdery mildew as the observation object, and probes into the rapid recognition method of wheat disease developed by suitable instrument in combination with image processing and pattern recognition technology, and designs a disease diagnosis system based on Android smart phone. The main content, innovation point and result are as follows: (1) The effects of high-pass filtering, middle value filtering and neighborhood averaging method are studied to reduce the influence of image acquisition environment; 3 segmentation algorithms (optimized watershed segmentation) are studied. Automatic threshold segmentation and horizontal set segmentation) to separate disease spots from wheat healthy leaves for extraction of disease spot characteristics. The disease characteristics of wheat leaf were described from three aspects: color, shape and texture (total extraction of 23 plaque characteristic parameters). The results show that the single image enhancement algorithm can not achieve the ideal enhancement effect, and the single image segmentation algorithm can not partition the target area well. Therefore, image enhancement and segmentation algorithms should be optimized to improve their enhancement and segmentation effects. (2) Three kinds of image recognition methods such as correlation vector machine (RVM), support vector machine (SVM) port inverse propagation neural network (BPNN) are studied. In this paper, 150 wheat leaf diseases (stripe rust and powdery mildew) of 150 different severity (including mild, moderate and severe) were selected as test materials. Among them, 68 disease leaves were selected as training samples, and the color of each disease blade was extracted. The texture and shape are 23 characters, and the weight of each feature in the disease color, texture and shape is calculated by using the Relief algorithm (i.e. the contribution size to the disease recognition), and 20 weight-weight features are selected as input parameters of the SVM, the BPNN and the RVM, and three identification models are respectively established. The results showed that the average recognition accuracy of SVM, BPNN and RVM was 86. 76%, 91. 17% and 89. 71%, respectively, while the accuracy of recognition of mild moderate disease was 86. 67%, 90. 00% and 88. 33%, respectively. The execution efficiency of RVM is 7.96 and 31.68 times that of SVM and BP neural network, respectively. (3) Aiming at the problems of inconvenience, high price, high professional requirement and so on, a diagnosis system of wheat leaf disease based on Android smart phone was developed in combination with RVM recognition algorithm. Sixty-six (33 powdery mildew and stripe rust) samples were collected by Sony DSC-H9 camera and SAMSUNG GT-N7100 cell phone, 48 of them (24 samples of powdery mildew and stripe rust) were selected as training samples and the rest were used as test samples. At the same time, the relationship between the pixel and the recognition rate is studied by changing the sampling pixel of the mobile phone as another control group, and the sample distribution is arranged on the same. The results show that the average recognition rate of RVM is 88. 89%, the correct recognition rate of disease is related to the acquisition tool and is directly proportional to its pixels. Therefore, it is necessary to select a suitable mobile phone for disease recognition to obtain higher accuracy. At the same time, through the application test, identification of a pair of disease pictures can be completed within 20s, and can realize rapid and accurate identification of the diseases of different severity leaf parts of wheat, which provides important technical support for the field investigation of plant protection personnel.
【学位授予单位】：安徽大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：S435.12;TP391.41

【参考文献】