复杂背景下马铃薯典型病虫害图像的分割与识别关键算法研究

The fast detection and recognition of diseases and pests is the basis of crop pests and diseases control. The rapid recognition and diagnosis of crop diseases and pests by computer vision and image recognition technology has become a hot research topic. The existed research mainly focuses on the recognition of diseases or pests in the simple background. However, there are few studies on the recognition of diseases and pests in the complex background, there are still many problems to be solved. In potato staple food of the strategic plan proposed by the Ministry of Agriculture, but also to accelerate the rapid recognition of potato diseases and pests. This project takes disease and pest of potato in the northern region of Yinshan mountain of Inner Mongolia as the research object, the acquisition and preprocessing methods of farmland disease and pest image under natural conditions were studied. The segmentation of disease and pest image was studied under complex background, in order to obtain accurate image of lesions and pests. The method of extracting and optimization of disease or pest color feature, texture feature and shape feature, in order to obtain the core features of the region. The method of image recognition based on pattern classification is studied in order to achieve the fast and accurate segmentation and recognition of diseases and pests. The research results will lay the foundation for intelligent recognition and diagnosis of crop diseases and pests in natural environment, and provide scientific theoretical basis for integrated prevention of crop diseases and pests in Inner Mongolia autonomous region.

病虫害快速检测与识别是农作物病虫害防治的基础，利用计算机视觉和图像识别技术进行农作物病虫害的快速识别与诊断已经成为研究热点，已有的研究主要集中于简单背景下的病害或虫害图像识别，而对复杂背景下病害和虫害图像的综合识别研究较少，还有诸多问题有待解决。在农业部提出的马铃薯主粮化战略布署下，更要加快马铃薯病虫害的快速识别研究。本项目以内蒙古阴山北麓旱作区马铃薯典型病虫害为研究对象，研究自然条件下大田病虫害图像的采集与预处理方法；研究复杂背景条件下病虫害图像的分割方法，以获得准确的病斑和害虫图像；研究病斑或害虫的颜色特征、纹理特征和形状特征的提取和优化方法，以得到能准确描述该区域的核心特征；研究基于模式分类的马铃薯病虫害图像识别方法，以达到对病虫害图像的快速、准确的分割与识别诊断。本项目的研究成果将为自然环境下农作物病虫害图像的智能识别诊断奠定基础，并为自治区农作物病虫害的综合防治提供科学理论依据。

如何准确快速地检测并识别诊断农作物病虫害，是有害生物防治的关键技术。本项目以自然条件下马铃薯病虫害图像为研究对象，运用图像处理与模式识别技术，实现了病虫害自动分类识别。项目研究实现了如何在复杂背景下准确分割出目标信息；根据病虫害视觉不变信息，提取并优化选择了对病虫害具有明显分类效果的特征，提出了小波域高频协方差阵特征值与低频低阶矩（HELM）纹理特征提取方法，可有效克服目标平移、旋转、缩放等问题，增强目标刻画能力；研究了多种特征融合机制，根据435维颜色、形状、纹理特征，提出了基于主成分分析的特征加权融合（FWFPCA）算法，不仅能提高特征表达水平，而且也可快速消除高维特征灾难，实现特征降维；研究建立了高效准确的分类器，针对13类病虫害融合特征，采用决策树思想，逐级进行支持向量机（SVM）分类，并以MATLAB R2012b为平台，对马铃薯典型的3类病害、10类虫害分别进行特征识别实验，验证了本项目提出算法的有效性。同时，项目选用6种典型马铃薯虫害作为研究对象，开发了一款采用深度学习算法嵌入移动端的马铃薯虫害识别系统，该系统采用迁移学习的方法，将虫害数据集训练生成的权重文件移植到移动端，在移动端中实现了图像的分类和结果显示。该系统识别准确率高、识别速度快、可操作性强，满足设计要求。本项目的研究成果将为自然环境下农作物病虫害图像的智能识别诊断奠定基础，并为自治区农作物病虫害的综合防治提供科学的理论依据。

内容获取失败，请点击重试