【作者】 李欣；

【导师】 张晋国； 张孟杰；

【作者基本信息】 河北农业大学 ， 农业机械化工程， 2007， 硕士

【摘要】 本文主要研究了田间杂草的图像识别技术，设计了田间杂草图像识别系统。系统包括麦田图像数据的采集，实现对图像的预处理；绿色植物与土壤背景的分割包括图像的灰度化与格式转换和图像的二值化；作物与杂草的分割包括作物中心行的识别和作物行的滤除，在滤除作物行的过程中确定边界阈值时本文提出了先计算标定的作物行宽度与计算机自动检测的作物行宽度之间的相对误差，然后选定合适的对应最小误差的作物行边界阈值的方法。识别杂草位置时分别采用了模板匹配和神经网络识别的方法。其中模板匹配识别杂草的具体方法是：从图像中选择一株草的特征数据作为模板，将模板一个象素一个象素的在图像中移动，即拿已知的模板和原图像中同样大小的一块区域去匹配，在搜索区域里寻找匹配点，以搜索窗口与目标物体形态特征的匹配度作为判据来实现目标检测与跟踪。本文提出采用计算模板象素与原图像中模板运行位置的象素之间的欧氏距离来验证它们之间的匹配度的方法，欧式距离越接近0，说明匹配程度越高。寻找到图像中与模板匹配度最高的点，识别为杂草，将其象素值标记为1，图像中显示为白色；在图像中以白色斑点的状态表示出来，其它象素作为背景标记为0，图像中显示为黑色。计算出白色斑点的坐标位置，即杂草的坐标位置，从而达到了寻找到杂草位置的目的。使用神经网络识别杂草的具体方法是：将使用3层完全结合方式的BP神经网络分类器对杂草图像进行分割，输入的特征量为每个象素的H,Cb,Cr值，所以输入层的神经元个数设计为3，输出层的神经元个数为1，输出0～1的信号，如果输出的信号大于0.9，相应的输入信号象素为杂草，将此象素标记为1，图像中显示为白色；如果输出信号小于0.1，相应的输入信号象素为土壤背景，将此象素标记为0，图像中显示为黑色。通过训练使网络达到稳定，最终使用MATLAB工具箱将图像分割，则图像中显示的白色斑点的位置即为杂草的位置。系统全程使用MATLAB语言编程。系统最终目的是根据杂草和作物分布的位置特征滤除作物行，识别出杂草，并计算出杂草的识别率。更多还原

【Abstract】 The identification technology of weed images in wheat fields was studied in this paper and a software system of weed images identification has been designed.The major content includes collecting images using digital camera; pretreating images; partitioning the green plants and the soil which includes again transforming the color images into grey images, and transforming the images’ formats, and transforming the grey images into binary images; partitioning the crop and the weed which includes again the identification of the center line of the crop and filtering of the crop line; finally obtaining the weed images. In order to compute the boundary of crop line, this paper choose different thresholds to compute the relative error between the automatically detected crop line width and the manually defined crop line width, and finally choose the appropriate threshold that minimum error.The template matching and the neural network recognition method are separately used to identify the position of weed.Template matching identification method is: Choosing a grass from the image to take the characteristic data of the template, moving a template picture from one element to another element in the image, Namely searching the region of the original map like the template and seek the match spots, and searching the window and characteristics of the goal object shape to realize the goal’s examination and the track as the criterion. computing Euclidean distance between the elements of template picture and the elements of the operating position of template in the original picture. If the Euclidean distance is closer 0, the match degree is higher. searching the highest matched degree spots. Weeds’pictures are composed by them. In the image its demonstrates white, and it express in the image by the white spots, and the pictures are composed by other elements. In the image its demonstrates black. If the weeds’ coordinated positions are computed, thus had achieved intention of seeking the weed position goal.The neural network identification method is: It use the BP neural network sorter of 3 tiers to car divide up for the weed image completely. The level unifies the way completely the BP neural network division sickness spot image. Input characteristics are each picture’s elements: H, Cb, Cr value, therefore the number of neuron center of input level is 3; the number of neuron center of Output level is 1. Then output signal from 0 to 1. If the output signal is bigger than 0.9, the image is weed. In the image it is demonstrated white; If the output signal is smaller than 0.1, the corresponding input signal picture element is the soil background, in the image it is demonstrated black. Enables the network through the training to achieve stably, finally uses the MATLAB toolbox to divide the image, then the image demonstrates white spots’ positions namely for weeds’ positions.This system was programmed by MATLAB language. So the system can successfully filter out the crop lines and identify the weeds’ positions.更多还原