基于BP神经网络的田间杂草识别技术的研究

发布时间：2018-12-13 05:36

【摘要】：农田伴生杂草的威胁是导致农作物生长势态萎靡、产量下降的主要原因,采用化学除草方式能高效及时地防治杂草幼苗,避免杂草成熟后影响农作物产量。化学除草多为大面积粗放式喷洒,简单省时却存有水土污染、农药残留、人员中毒等诸多弊端,与当下提倡的绿色环保、可持续发展、精准农业等理念相悖,加之现代农作物的播种多为定点条播式,杂草以簇生且随机性生长在垄间,因此需要设计一种定点变量式喷药系统。本研究以吉林农业大学玉米试验田间常见杂草的防治问题为例,设计了以机器视觉和图像处理技术相结合的杂草识别系统,为实时除草设备的研发提供了基础条件。本文的主要研究内容如下:1.分析比较杂草与作物的出苗速度以及后续处理对图像质量的要求,从而确定图像采集的时间和采集过程镜头的高度与角度;提取图像在RGB、2G-R-B、HIS、YCbCr等颜色空间内的颜色特征,确定2G-R-B特征值最能满足灰度化的要求;分别利用邻域均值滤波和中值滤波两种算法消除图像采集、传输和变换的过程中引入的噪声和干扰,比较处理结果发现中值滤波对带有椒盐噪声的田间图像的处理效果更好。2.列举三种阈值分割方法进行比较分割,其中OTUS阈值分割法的运行速度最快,前景植物图像完整且没有噪声干扰,但背景区域噪声较多;利用数学形态学开运算进行后处理,比较不同尺寸的腐蚀单元和膨胀单元的作用效果,最终选定半径为23的平面圆盘腐蚀算子和直径为13的棱形膨胀算子。3.对形态学分割完成后的独立叶片进行连通区域标记,计算基于区域特征的面积和矩特征等参数;利用5种边缘检测方法对标记后的连通区域进行边缘检测,发现Canny算子检测结果最为理想,然后基于轮廓特征计算各区域的周长、长和宽等有量纲参数;对比玉米、稗草和苋菜三种植物的特征参数发现宽长比、圆形度和第一不变矩能有效区分植物的种类,宜作为杂草分类器的输入特征参数。4.人工神经网络对于复杂的田间图像具有独特的存储、联想和识别判断的功能,因此建立基于BP神经网络的杂草分类器;对隐层节点数、学习率和动量因子进行试验设计分析得到了最优组合解,确定了网络结构为MLP:361,学习率为0.5,动量因子为0.5;利用MATLAB仿真建立BP神经网络,并以田间作物和杂草共120幅图像的特征参数作为输入进行识别分析,其中90组特征参数作为训练样本,30组作为测试样本,仿真得到的识别正确率分别为98.89%和93.33%。
[Abstract]:The threat of associated weeds in farmland is the main reason leading to the decline of crop growth and yield. The chemical weeding method can effectively and timely control weed seedlings and avoid the effect of weeds ripening on crop yield. Chemical weeding is mostly extensive spraying in a large area. Simple and time-saving, however, there are many drawbacks such as soil and water pollution, pesticide residues, personnel poisoning, etc., which are contrary to the current concepts of green environmental protection, sustainable development, precision agriculture, etc. In addition, the sowing of modern crops is mostly fixed spot seeding, weeds are clustered and growing randomly between ridges, so it is necessary to design a fixed point variable spraying system. Taking the common weed control problems in maize experimental field of Jilin Agricultural University as an example, a weed recognition system based on machine vision and image processing is designed, which provides a basic condition for the research and development of real time weeding equipment. The main contents of this paper are as follows: 1. The emergence speed of weeds and crops and the requirements of image quality for subsequent processing were analyzed and compared to determine the time of image acquisition and the height and angle of lens during acquisition. Extracting the color feature of the image in the color space such as RGB,2G-R-B,HIS,YCbCr and determining the 2G-R-B feature value can meet the requirement of grayscale most. The neighborhood mean filter and median filter are used to eliminate the noise and interference in the process of image acquisition, transmission and transformation, respectively. Compared with the processing results, the median filter has better effect on the field image with salt and pepper noise. 2. 2. Three threshold segmentation methods are listed. Among them, OTUS threshold segmentation method has the fastest running speed, the foreground plant image is complete and has no noise interference, but the background region noise is more. The effect of corrosion unit and expansion unit with different sizes were compared by using mathematical morphological open operation. Finally, the plane disk corrosion operator with radius 23 and the prism expansion operator with diameter 13 were selected. The independent leaves after morphological segmentation were labeled with connected region, and the parameters such as area and moment characteristics based on regional features were calculated. Five edge detection methods are used to detect the marked connected regions, and the results of Canny operator detection are found to be the most ideal. Then the dimensionality parameters such as perimeter, length and width of each region are calculated based on the contour features. Comparing the characteristic parameters of corn, barnyard grass and amaranth, it was found that the ratio of width to length, roundness and the first invariant moment could effectively distinguish the species of plants, and it should be used as the input characteristic parameter of weed classifier. 4. Artificial neural network (Ann) has unique functions of storing, associating and judging complex field images, so a weed classifier based on BP neural network is established. The optimal combination solution is obtained by analyzing the number of hidden nodes, learning rate and momentum factor. The network structure is determined as MLP:361, learning rate of 0.5 and momentum factor of 0.5. The BP neural network was established by MATLAB simulation, and the characteristic parameters of 120 images of field crops and weeds were used as input for identification and analysis. 90 groups of characteristic parameters were used as training samples, 30 groups as test samples. The recognition accuracy obtained by simulation is 98.89% and 93.33% respectively.
【学位授予单位】：吉林农业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：S451;TP391.41

【参考文献】