【作者】 吴翔；

【导师】 裘正军；

【作者基本信息】 浙江大学 ， 生物系统工程， 2016， 博士

【摘要】 准确识别害虫是虫情监测防治的重要前提。目前我国农作物的害虫诊断主要依靠有限的植保人员及昆虫专家进行人工识别,费时费力且覆盖面有限。基于机器视觉的害虫识别方法,能够解决人工识别中存在的人力不足、识别率低等问题,可以实现及时有效的防治,具有很高的研究价值和意义。由于害虫个体小、种类繁多等原因,害虫的识别问题相比其他机器视觉应用具有更大的实现难度。本文以典型农业害虫作为试验对象,研究了基于图像特征提取、卷积神经网络(CNN)和高光谱成像技术的害虫识别方法,并设计了基于图像的害虫识别系统,主要内容有：(1)提出了基于图像特征提取的害虫识别方法。主要采用“图像分割——特征提取——分类器设计”的技术路线,在图像分割阶段,利用角点检测算法对原始图像进行裁剪以减少背景区域,改善了在不均匀自然光、目标较小的情况下Otsu法难以计算最佳阈值的缺陷。针对图像分割后的害虫目标,提取了形态、颜色和纹理等全局特征,同时利用SURF算法提取害虫的局部特征。分别基于全局特征和局部特征建立SVM分类模型。对9类害虫进行识别,全局特征分类模型的总体识别率为85.9%,局部特征分类模型的总体识别率为77.4%。(2)提出了基于CNN的自然背景下害虫识别方法。根据害虫图像集的特点,构建一个12层的CNN模型,在CNN基本结构的基础上添加了归一化层,使用不饱和函数ReLu作为激励函数,并根据开发平台的特性对其他功能层进行优化调整。该模型以128×128的彩色图像作为输入,经过隐藏层的映射变换,输出对应的类别值。对10类害虫图像进行测试,当训练次数达到45次时,模型的测试集识别率为76.7%。进一步研究了图像预分割对CNN模型的影响,通过GrabCut预分割来降低图像集的识别难度,从而提高CNN模型的识别效果。(3)研究了基于高光谱成像技术的害虫检测和识别方法。采集菜青虫和包菜叶片的近红外高光谱数据(1000nm-1600nm),基于特征波长建立PLS-DA和BPNN模型,在预测集上均达到了96%以上的准确率。选择SPA-PLS-DA模型,结合图像处理算法,对菜青虫混合样本的高光谱图像进行可视化分析,以二值图像的形式显示包菜叶片上菜青虫的空间位置和虫体形状,实现了菜青虫的可视化检测。采集稻纵卷叶螟、二化螟、黄杨绢野螟、玉米螟4种螟蛾科害虫的近红外高光谱数据,对比了原始光谱数据(Raw)和3种光谱预处理方法(SG、MSC、SNV),发现原始光谱数据为最优。基于SPA算法选取9个特征波长,分别建立PLS-DA、BPNN、ELM和SVM识别模型,结果表明ELM模型取得最好的识别效果,建模集和预测集的准确率均为100%。(4)设计了基于图像的害虫识别系统。在图像采集方面,使用工业摄像头和可调节支架搭建本地采集平台,通过网络摄像头建立了害虫图像的远程采集和传输系统。开发了基于图像特征提取的害虫识别软件,包括图像分割、特征提取、模型训练等功能模块,能够对本地硬盘图像和摄像头采集的实时图像进行识别。开发了基于CNN的害虫识别软件,实现CNN模型的训练与识别功能,具有重新训练和继续训练两种训练方式,利用保存的模型对害虫图像进行识别。更多还原

【Abstract】 Precise identification of pests is a requisite for integrated pest management (IPM). The traditional method of pest identification is mainly achieved by limited plant protection personnels and entomologists, which is laborious and time-consuming. Machine vision can solve the problems of labor shortage and low accuracy in artificial identification, which has high research value. Compared to other machine vision application, pest identification is more difficult due to small size of pests and wider varieties. In this context, identification of agricultural typical pests was studied, based on feature extraction, convolutional neural network (CNN) and hyperspectral imaging. Besides, pest image indetificaition systems were developed. The main contents of the thesis are as follows:(1) The pest identification method based on image feature extraction was proposed, with a route of "image segmentation--feature extraction--classifier design". In the phase of image segmentaion, a cornor detection algorithm was applied for image cropping to reduce the background area, which could improve the ability of Otsu segmentation method for finding the optimal threshold under the condition of uneven ambient light or smaller object. Based on the segmented pest target, global features were extracted, including color, morphological and texture characteristics. Besides, the SURF algorithm was explored to extract local features, which was invariant to image scaling, translation and rotation. The SVM classifier was used to identify 9 kinds of pests. The accuracy using global features achieved 85.9% and the accuracy using local features was 77.4%.(2) In order to realize recognition of pest images under natural backgrounds, a method based on CNN was proposed. Considering the traits of the pest image set, a CNN model of 12 layers was constructed. On the basis of the basic structure of CNN, normalization layer was added to acquire better generalization and Relu was applied as excite function. Other hidden layers were also adjusted according to the development platform. The CNN model used RGB images of 128*128 as input, after mapping transform of hidden layer, and then calculated the corresponding class values. When training epoch reached 45, the recognition rates of test set was 76.7%. Besides, the effect of image pre-segmentation on the CNN model was explored. The result showed that the GrabCut segmentation method could reduce the recogntion difficulty of the original pest image set, leading to higher model accuracy.(3) Hyperspctral imaging technolodgy was applied for detection and recogntion of pests. Two objectives were studied using NIR wavelength(1000nm-1600nm)。 The first objective was detection of Pieris rapae larvae:The developed SPA-PLS-DA model, combined with image process algorithm, was used for pixel-wise detection of larvae in the hyperspectral images of mixed samples; a binary image was produced to display the location and approximate shape of larvae precisely and intuitively. The second objective was recognition of pests:hyperspectral data of Cnaphalocrocis medinalis Guenee, Chilo suppressalis, Diaphania perspectalis and Ostrinia nubilalis were acquired; different preprocessing methods (Raw, SG, SNV and MSC) were implemented, and the result showed that the Raw was the best. Based on the 9 EWs selected by SPA,4 models of PLS-DA, BPNN, ELM, SVM were developed; by comparison, the SPA-ELM model achieved the best prediction result, with 100% classification accuracy for both the calibration set and the prediction set.(4) Pest identification systems were designed. An industrial camera was used to set up an indoor image acquisition platform. And a webcam was used for remote acquisition and transmission of pest images. Pest identification software based on image feature extraction was developed, which had functionalities of image segmentation, feature extraction. This software could identify the pest image acquired from local disk and usb camera. Pest identification software based on CNN was developd, which realized training and recognition of CNN model. This software had two training modes, and used saved model to indentify pest images.更多还原