基于图像的稻田飞虱测报方法的研究

发布时间：2018-04-01 09:06

本文选题：白背飞虱测报　切入点：图像处理　出处：《浙江理工大学》2017年硕士论文

【摘要】：稻飞虱是我国水稻上最重要的一类远距离迁飞性害虫,各稻区以褐飞虱,白背飞虱和灰飞虱最为常见,它们常群居在水稻中下部取食或产卵为害。稻飞虱不同虫态的田间发生调查是其进行准确测报和合理防治的关键。在大部分水稻种植国家,稻飞虱田间调查一般采用人工拍查法、目测法、扫网法和灯诱法,其中,应用最为广泛是拍查法。“拍查法”查获的飞虱率受到虫口密度、水稻生育期、盘内壁湿润程度等影响且容易造成调查者身体和视觉疲劳,调查效率低下。人工田间调查时只记录了稻飞虱种类和各种虫态数量的数据,后面无法追溯田间调查时飞虱发生的真实情况。刘庆杰等[1-4]充分利用数字图像处理技术,研究了不同特征对水稻基部稻田飞虱检测的影响,取得了较好的检测效果,但飞虱误检率仍旧偏高。本文在此基础上,采用新的三层检测方法,研究了图像特征和分类模型参数的选择对水稻基部飞虱检测率和误检率的影响,以及白背飞虱各虫态(包括长翅型成虫、短翅型成虫、高龄若虫和低龄若虫)的分类问题。本论文主要研究内容、研究结果和创新点包括:(1)在第一层检测中,研究了不同维数的HOG特征和不同级联层数的Adaboost分类器对水稻基部白背飞虱检测率和误检率的影响。首先,对2012-2016年期间采集的水稻基部飞虱图像建立了白背飞虱和非飞虱噪声的正负训练样本集;然后,提取训练样本不同维数的HOG特征;利用HOG特征训练不同级联层数的Adaboost分类器,用于检测稻飞虱;最终选择最优Adaboost分类器测试525张水稻基部飞虱图像。结果表明该算法对白背飞虱检测率为90.7%,误检率为56.2%。(2)在第二层检测中,针对第一层中存在较多的误检噪声,研究了不同局部图像特征训练的SVM分类器对非飞虱噪声识别情况。这些噪声主要包括水珠、水面反光、泥点以及稻叶,它们在纹理上与飞虱存在较大差异。首先提取训练样本的Gabor与LBP特征,使用Z-score进行归一化;然后,利用Gabor、LBP和两个特征融合来训练SVM分类器,根据不同特征训练获得的SVM分类器ROC曲线,发现Gabor和LBP融合的纹理特征训练的SVM分类器对白背飞虱和非飞虱噪声识别率高;最终利用该SVM分类器对525张第一层检测后得到的子图像进行非飞虱噪声排除;结果表明该算法将第一层的误检率从56.2%降低到了10.2%。(3)在第三层检测中,研究了白背飞虱不同虫态分类识别的问题。针对不同虫态的白背飞虱HOG特征差异显著,本文首先提取了白背飞虱三种虫态长翅型成虫、高龄若虫和低龄若虫HOG特征,并使用Z-score进行归一化;然后,利用PCA与LDA方法对HOG特征进行降维,比较不同降维算法对白背飞虱不同虫态的识别性能;最后,采用SVM分类器对525张水稻基部飞虱图像第二层检测到的白背飞虱子图像进行虫态分类识别;结果表明该算法对白背飞虱长翅型成虫、高龄若虫和低龄若虫识别率分别为93.2%、82.7%和86.9%。综合三层检测结果,最终获得水稻基部白背飞虱各虫态平均识别率为73.1%。误检率为23.3%。对于无虫的图像,误检率为5.6%。由此可见,利用图像处理方法进行水稻基部飞虱测报是可行的。
[Abstract]:Rice is the most important rice in China on a long-distance migratory insect, the rice to brown planthopper, sogatellafurcifera and l.striatellus was the most common, they often populations in Rice under feeding or oviposition. Occurrence of rice planthopper infestation of different insect state investigation is the key to accurate forecasting and rational prevention in most countries. Planting rice, rice field investigation using artificial shoot check method, visual method, sweep net method and light trap method, among them, the most widely used is to take check method. "Shoot check method" seized by planthopper insect density, rice growing period, moist degree and the influence of wall plate easy to cause the body survey and visual fatigue investigation efficiency. Artificial field investigation only records the number of rice planthopper types and various stages of data, the real situation cannot be traced back behind the field investigation. Liu Qingjie planthopper occurrence [1-4] charge Using digital image processing technology, the effects of different characteristics on the base of rice brown planthopper in rice field detection, a good detection effect, but the planthopper false detection rate is still high. On this basis, using three layers of new detection methods, study the image features and classification of the model parameters on the effect of rice base planthopper detection rate and false detection rate, and the white backed planthopper of different stages (including the macropterous and brachypterous adults, aged nymphs and young nymphs) classification problems. The main contents of this dissertation, the research results and innovations include: (1) in the first layer on the HOG feature detection. Different dimensions and the associated number of Adaboost classifier on the white back planthopper detection rate and false detection rate. First of all, on the base of the rice planthopper image acquisition period of 2012-2016 years, established the white backed Planthopper and the noise is non planthopper The negative training set; then, HOG features are extracted from different dimensions of training samples; Adaboost classifier based on HOG features of different layers of the cascade training, for the detection of rice planthopper; finally choose optimal Adaboost classifier test 525 rice base planthopper images. The results show that the algorithm WBPH detection rate is 90.7%, the error rate is 56.2%. (2 in the second layer) detection, the false noise exists in the first layer, the SVM classifier trained on different local image features non noise recognition. These noise planthopper mainly include water, water surface reflection, mud and rice leaves, they in the texture and planthopper first extract the Gabor there is a big difference. With the LBP feature of training samples, then using Z-score normalization; using Gabor, LBP and two features to train SVM classifier, SVM classifier based on ROC features of the different training The curve, found that Gabor and LBP fusion texture feature training SVM classifier WBPH and non noise planthopper high recognition rate; finally uses the SVM method to get the first 525 layers after detection of sub images are non noise planthopper exclusion; the results show that the algorithm will first layer error rate is reduced from 56.2% to 10.2%. (3) in the third layer detection, study the WBPH different insect state classification problems. According to the difference of HOG features sogatellafurcifera different stages significantly, the paper extracts sogatellafurcifera three instars macropterous nymphs, elderly and young nymph HOG characteristics, and normalized using Z-score; then, to reduce the dimensionality of HOG features using PCA and LDA method, the recognition performance comparison of different dimensionality reduction algorithm of WBPH at different developmental stages; finally, using SVM method to detect 525 rice planthopper images of second Zhang base layer to the White - backed image classification and recognition of lice insect state; the results show that the algorithm to WBPH macropterous adults, aged nymphs and nymph recognition rates were 93.2%, 82.7% and three layer 86.9%. results comprehensive detection, finally get the white back planthopper insect the average recognition rate is 73.1%. error rate is 23.3%. for no image of the insect, the false detection rate is 5.6%. thus, using the image processing method of rice planthopper at base is feasible.

【学位授予单位】：浙江理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：S435.112.3;TP391.41

【参考文献】