中国农业科学 ›› 2014, Vol. 47 ›› Issue (4): 664-674.doi: 10.3864/j.issn.0578-1752.2014.04.006

• 耕作栽培·生理生化·农业信息技术 • 上一篇    下一篇

基于计算机视觉的水稻叶部病害识别研究

 刘涛1, 仲晓春2, 孙成明1, 郭文善1, 陈瑛瑛1, 孙娟1   

  1. 1、扬州大学农学院/江苏省作物遗传生理国家重点实验室培育点,江苏扬州 225009;
    2、中国农业科学院农业信息研究所,北京 100081
  • 收稿日期:2013-06-13 出版日期:2014-02-15 发布日期:2013-09-22
  • 通讯作者: 孙成明,E-mail:cmsun@yzu.edu.cn;郭文善,E-mail:guows@yzu.edu.cn
  • 作者简介:刘涛,E-mail:tliu823@163.com
  • 基金资助:

    江苏高校优势学科建设工程资助项目(2011-05)、江苏省农业三新工程资助项目(SXGC[2012]399)

Recognition of Rice Leaf Diseases Based on Computer Vision

 LIU  Tao-1, ZHONG  Xiao-Chun-2, SUN  Cheng-Ming-1, GUO  Wen-Shan-1, CHEN  Ying-Ying-1, SUN  Juan-1   

  1. 1、College of Agriculture, Yangzhou University/Key Laboratory of Crop Genetics and Physiology of Jiangsu Province, Yangzhou 225009, Jiangsu;
    2、Institute of Agriculture Information, Chinese Academy of Agriculture Sciences, Beijing 100081
  • Received:2013-06-13 Online:2014-02-15 Published:2013-09-22

摘要: 【目的】文章重点分析了病健交界特征参数、病害识别流程对提高病害识别准确率的影响。实现水稻叶部15种主要病害的准确识别,尤其是相似病害的判断。【方法】(1)病斑图像获取:水稻叶部病害图像来源包括水稻大田、病害图册和病害数据库,文中选用改进的mean shift图像分割算法提取病叶图像中的病斑并根据相关方程获取病斑特征信息。(2)特征参数的选择与设计:首先选取一至三阶颜色矩和颜色直方图作为病害的颜色特征参数,选取球状性、偏心率和不变矩作为病斑的形状特征参数,选取角二阶矩、对比度和相关作为病斑的纹理特征参数;然后针对相似病斑误报率高的问题提出一种病健交界特征参数,通过病斑内部、边缘和外围颜色上的差异描述该特征,并根据3个区域相互间归一化颜色直方图的欧氏距离计算该项特征参数,该参数可以用于描述病斑与健康部分交界处的特征。(3)病害识别流程的设计:根据病害在颜色、形状、纹理、病健交界4个特征上差异的显著程度设计完成病害识别流程,流程中首先通过颜色特征识别病害,对于通过颜色特征无法识别的病害再通过形态特征识别,倘若形态特征依然无法识别则通过纹理和病健交界特征进行最终识别。(4)病害识别模型的建立:将病害数据分成两部分,一部分用于建立模型,另一部分用于模型的验证;利用LibSVM程序包完成建模,其中svmtrain函数用于建立支持向量机模型,Grid程序用于优化参数,svmpredict函数用于对模型进行验证。【结果】15种水稻叶部病斑可以从复杂的背景中分割出来,并可快速准确的被识别,平均识别准确率为92.67%,平均漏报率为7.00%,最大漏报率和误报率分别为15.00%和25.00%;病健交界特征参数引入后,识别准确率提高了14.00%,平均漏报率降低了7.50%,漏报率最大降幅为20.00%,误报率最大降幅为65.00%;与用所有特征参数直接进行病害识别相比,采用本文提出的识别流程进行病害识别的准确率提高了12.67%,漏报率降低了9.33%,一些病害的漏报率和误报率降幅达30.00%以上;在识别流程各步骤中,颜色特征识别环节的平均准确率为96.71%,漏报率和误报率均未超过10.00%;形态特征识别环节的平均准确率为94.17%,漏报率和误报率均未超过15.00%;纹理和病健交界特征识别环节的平均准确率为91.50%,漏报率和误报率均未超过25.00%。【结论】利用mean shift图像分割算法可以准确分割水稻叶部病斑;基于支持向量机模型的分类方法可以对15种水稻病斑准确分类;论文中提出的病健交界特征参数以及病斑识别流程均提高了病斑的识别准确率;病健交界特征参数对提高一些相似病害的识别精度效果显著;将这些方法相结合可以有效对水稻常见叶部病害进行识别,为水稻病害的田间智能诊断提供技术支撑。

关键词: 水稻病害 , 计算机视觉 , 图像处理 , 病健交界 , 均值漂移 , 支持向量机

Abstract: 【Objective】The purpose of this article was to recognize 15 kinds of rice disease located at leaf accurately, especially among the similar diseases. 【Method】 For disease spot image acquisition, the original image was got by digital camera from paddy fields and modified mean shift image segmentation algorithm was used to extract rice disease spot from the original image. For the selection and design of feature parameters, Firstly, the color moments arranged from first to third order and color histogram were selected as color feature parameters, and sphericity, eccentricity, invariant moment as shape feature parameters, and angular second moment, contrast, correlation as texture feature parameters. Secondly, junction feature parameters were designed aiming at high missing report rate of similar diseases, this feature was established based on the difference between inside, margin and periphery of disease spot, and euclidean distances of color in different regions were used to calculate this feature parameters which could describe the junction of health and disease. Thirdly, for the design of diseases identification process, the process of diseases was designed based on the key characters of each disease, in this process, diseases were distinguished by the biggest character difference. The concrete step of identification process are as follows: Firstly, color feature was used to divide all the diseases into six groups. Secondly, shape feature was used to subdivide. Finally, texture and junction feature were used for ultimate recognition. Color, shape, texture and junction feature parameters were obtained by related equation, and the parameters were divided into fifteen groups by their disease types. Fourthly, for the establishment of recognition model, the Support Vector Machine(SVM) model was applied to classify and recognize the 15 kinds of rice diseases stepwise. The images were divided into two groups, one was selected to build the model, and another to verify it. LibSVM software package was used to modeling, of which svmtrain function, svmpredict function and grid software were separately used to establish and verify the model and parameter optimization. In this research, a comprehensive study was conducted about the impact of junction feature parameters and identification process on improving accuracy.【Result】A total of 15 kinds of rice disease could be extracted and recognized by this method successfully, and the average recognition accuracy rate was up to 92.67% , average missing report rate was 7.00%, the max missing report rate and false report rate was 15.00% and 25.00%, respectively. After using junction feature parameters, recognition accuracy rate was 14.00% higher, average missing report rate was 7.50% lower, the max decreasing ranges of missing report rate and false report rate were 20.00% and 65.00%, respectively. The identification process proposed in this study could make the recognition accuracy rate 12.67% higher, missing report rate was 9.33% lower, and decreasing ranges of some diseases were more than 30.00%. In all the steps of diseases identification process, the accuracy rate of color recognition was 96.71%, missing report rate and false report rate were both under 10.00%. At the second step, the accuracy rate was 94.17%, missing report rate and false report rate were both under 15.00%. The accuracy rate of texture and junction feature step recognition was 91.5%, missing report rate and false report rate were both under 25.00%.【Conclusion】The 15 kinds of rice disease spots could be segmented from the original images based on the modified mean shift image segmentation algorithm, and the disease could be accurately classified with the method of SVM model with proper feature parameters. The junction feature parameter and the classifying process which were put forward in this paper could improve recognition accuracy. The junction feature parameter could increase the recognition accuracy significantly, especially when the target diseases are similar. Common rice leaf diseases could be recognized by using all those technologies. This paper offered a technical support for automatic diagnoses of rice diseases further.

Key words: rice disease , computer vision , image processing , junction of health and disease , mean shift , support vector machine