早期玉米苗与杂草的自动辨识算法研究
发布时间：2018-08-20 14:48
【摘要】：随着计算机技术的飞速发展,在计算机视觉领域内的图像处理和模式识别等技术也逐渐变得更加完善,其中检测问题也是图像处理等领域大多数学者研究的主要问题之一,并且在实际生活中,检测问题也有着十分广泛的实际应用。在农业生产领域中,谷类作物是我国粮食十分重要的来源之一,而玉米作为主要的谷类作物,玉米早期的幼苗能够健康大量的生长对我国粮食安全以及工业生产都起到非常不可小视的作用。所以能在玉米生长的初期准确高效的去除杂草是非常必要的。目前为止人们普遍使用传统的除草方式,例如人工除草和除草剂除草等等。这些传统的方法虽然可能会有很高的准确率,但是会消耗很大的人力,这就无形中提高了人工成本。并且如果大量喷洒化学除草剂,不仅会对食用该谷物的人的健康造成威胁,也会对环境产生严重的污染。同时如果长时间的利用除草剂进行除草,也逐渐会使土壤对于除草剂有较强的依赖性,因此除草剂也不是持续性的除草方案。所以找到一种快速便捷的除草方法是十分必要的。基于上述提出的问题,本文从计算机视觉以及深度学习中领域出发,以能够快速有效的辨识玉米幼苗与杂草为目标。旨在分析和探索能够运用计算机视觉领域内的知识来解决自动辨识玉米的方法。本文运用了计算机视觉领域中处理检测问题的相关知识,提出了能够快速辨识玉米幼苗与杂草的方法,并通过在自己的数据集中进行了大量实验,对提出的算法是否可行进行论证。本文首先分析了传统除草方法存在的一些缺点,以及智能除草对于精确度以及处理速度有着较高的要求,之后通过在温室中拍摄大量的玉米幼苗与杂草的图片来构成数据集。通过观察玉米幼苗与杂草的主要区别以及联想计算机视觉领域检测问题的一些处理方法,首先通过对所采集到的数据集进行预处理工作,去除光照以及噪声的影响,之后分别采用两种方向来实现早期玉米苗与杂草的自动辨识。首先本文采取传统的人工手动选取特征的方法,通过观察早期玉米幼苗与杂草的主要区别来相应的选择特征。并且根据不同的特征的特点选取两种特征分别对样本进行特征提取,之后将两种特征点进行融合,提取特征向量并且使用分类器训练提取到的特征向量,最终得到可以区分玉米幼苗与杂草的分类模型。第二种方法借助深度学习中卷积神经网络可以分类的特点,基于目前比较流行的Faster R-CNN检测模型,借助区域建议网络RPN以及用于分类的Fast R-CNN检测器,通过对自己采集的数据集进行人工标注,调整网络的结构和参数,训练自己的数据集,最终得到可以用于分类的模型,实现早期幼苗与杂草的分类。整个过程我们从实际问题需要出发。利用计算机视觉领域的知识为依据,提出了解决问题的方法,实现了早期幼苗与杂草的自动辨识。最后,我们对所有的论文的内容进行了概括,提出目前所做工作需要改进的地方与此同时指出将来需要研究的内容。
[Abstract]:With the rapid development of computer technology, image processing and pattern recognition technology in the field of computer vision are becoming more and more perfect. Detection problem is also one of the main problems studied by most scholars in the field of image processing, and in real life, detection problem has a very wide range of practical applications in agriculture. In the field of industrial production, cereal crops are one of the most important sources of grain in China. As a major cereal crop, maize seedlings can grow healthily and massively in the early stage, which plays a very important role in food security and industrial production in China. It's necessary. So far traditional weeding methods, such as artificial weeding and herbicide weeding, are widely used. Although these traditional methods may have high accuracy, they will consume a lot of manpower, which will invisibly increase the cost of labor. And if a large number of chemical herbicides are sprayed, not only will they be used for food. At the same time, if long-term use of herbicides for weeding, the soil will gradually become more dependent on herbicides, so herbicides are not a sustainable weeding program. Therefore, it is necessary to find a fast and convenient weeding method. In order to solve the above problems, this paper starts from the fields of computer vision and in-depth learning, aiming at identifying maize seedlings and weeds quickly and effectively. This paper presents a method to identify maize seedlings and weeds quickly, and demonstrates the feasibility of the proposed algorithm through a large number of experiments in our own data set. Firstly, this paper analyzes some shortcomings of traditional weeding methods, and intelligent weeding has a higher precision and processing speed. High requirements, then by taking a large number of images of maize seedlings and weeds in the greenhouse to form a data set. First of all, the traditional manual feature selection method is adopted to select the characteristics of early maize seedlings and weeds by observing the main differences between early maize seedlings and weeds. After feature extraction, the two feature points are fused to extract feature vectors and trained by classifiers. Finally, a classification model which can distinguish maize seedlings from weeds is obtained. The second method uses convolution neural network in depth learning to classify, which is based on the popular Faster. R-CNN detection model, with the help of RPN and Fast R-CNN detector for classification, labels the collected data sets manually, adjusts the structure and parameters of the network, trains the data sets, and finally gets the model that can be used to classify the early seedlings and weeds. Based on the knowledge of computer vision, this paper presents a method to solve the problem and realizes the automatic identification of early seedlings and weeds. Finally, we summarize the contents of all the papers, and point out what needs to be improved and what needs to be studied in the future.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41
本文编号：2194022
[Abstract]:With the rapid development of computer technology, image processing and pattern recognition technology in the field of computer vision are becoming more and more perfect. Detection problem is also one of the main problems studied by most scholars in the field of image processing, and in real life, detection problem has a very wide range of practical applications in agriculture. In the field of industrial production, cereal crops are one of the most important sources of grain in China. As a major cereal crop, maize seedlings can grow healthily and massively in the early stage, which plays a very important role in food security and industrial production in China. It's necessary. So far traditional weeding methods, such as artificial weeding and herbicide weeding, are widely used. Although these traditional methods may have high accuracy, they will consume a lot of manpower, which will invisibly increase the cost of labor. And if a large number of chemical herbicides are sprayed, not only will they be used for food. At the same time, if long-term use of herbicides for weeding, the soil will gradually become more dependent on herbicides, so herbicides are not a sustainable weeding program. Therefore, it is necessary to find a fast and convenient weeding method. In order to solve the above problems, this paper starts from the fields of computer vision and in-depth learning, aiming at identifying maize seedlings and weeds quickly and effectively. This paper presents a method to identify maize seedlings and weeds quickly, and demonstrates the feasibility of the proposed algorithm through a large number of experiments in our own data set. Firstly, this paper analyzes some shortcomings of traditional weeding methods, and intelligent weeding has a higher precision and processing speed. High requirements, then by taking a large number of images of maize seedlings and weeds in the greenhouse to form a data set. First of all, the traditional manual feature selection method is adopted to select the characteristics of early maize seedlings and weeds by observing the main differences between early maize seedlings and weeds. After feature extraction, the two feature points are fused to extract feature vectors and trained by classifiers. Finally, a classification model which can distinguish maize seedlings from weeds is obtained. The second method uses convolution neural network in depth learning to classify, which is based on the popular Faster. R-CNN detection model, with the help of RPN and Fast R-CNN detector for classification, labels the collected data sets manually, adjusts the structure and parameters of the network, trains the data sets, and finally gets the model that can be used to classify the early seedlings and weeds. Based on the knowledge of computer vision, this paper presents a method to solve the problem and realizes the automatic identification of early seedlings and weeds. Finally, we summarize the contents of all the papers, and point out what needs to be improved and what needs to be studied in the future.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41
【参考文献】
相关期刊论文 前4条
1 阎庆;梁栋;张晶晶;;基于Fisher变换的植物叶片图像识别监督LLE算法[J];农业机械学报;2012年09期
2 李慧;祁力钧;张建华;冀荣华;;基于PCA-SVM的棉花出苗期杂草类型识别[J];农业机械学报;2012年09期
3 毛文华;曹晶晶;姜红花;王一鸣;张小超;;基于多特征的田间杂草识别方法[J];农业工程学报;2007年11期
4 傅弘,池哲儒,常杰,傅承新;基于人工神经网络的叶脉信息提取——植物活体机器识别研究Ⅰ[J];植物学通报;2004年04期
相关博士学位论文 前1条
1 吴清锋;基于内容的中草药植物图像检索关键技术研究[D];厦门大学;2007年
，本文编号：2194022
本文链接：https://www.wllwen.com/kejilunwen/ruanjiangongchenglunwen/2194022.html
