【作者】 马本学；

【导师】 应义斌；

【作者基本信息】 浙江大学 ， 生物系统工程， 2009， 博士

【摘要】 针对目前库尔勒香梨检测分级手段落后,检测效率低,而且对其内部品质检测只能进行有损检测。综合利用图像处理技术、光谱分析技术、高光谱成像技术、模式识别、化学计量学、光学、果树生理学等诸多领域的知识,开展对其外部品质和内部品质（类别、重量、糖度、果梗,以及形状等）的无损检测方法研究,在上述研究的基础上建立其品质的快速检测方法和检测模型,为以后进一步研究开发库尔勒香梨品质在线无损检测装备提供理论支持,且得出的研究结论可借鉴用于其它同类水果。主要研究内容和结论为:（1）基于图像处理技术的库尔勒香梨脱萼果和宿萼果判别研究。依据图像处理技术提取香梨底部的花萼特征,计算出其圆度,然后根据统计特性确定阈值对脱萼果和宿萼果进行判别分析,研究结果表明该方法对脱萼果的正确识别率为79.7%,对宿萼果的正确识别率为85.9%,对所有香梨样本的正确识别率为82.8%。结果表明,利用图像处理技术能对库尔勒香梨脱萼果和宿萼果进行自动判别。（2）基于近红外光谱分析技术的库尔勒香梨脱萼果和宿萼果判别研究。对比分析了应用DA判别分析和SIMCA法在不同波段、不同预处理方法对库尔勒香梨脱萼果和宿萼果的定性分析模型性能的影响。结果表明:利用近红外光谱分析技术,判别分析（Discriminant Analysis,DA）法在长波近红外区域(9091～4000 cm^-1)范围结合原始光谱建立的DA判别模型最优,对校正集正确分类率为100%,预测集正确分类率为95%;而采用SIMCA法利用近红外漫反射光谱技术对库尔勒香梨脱萼果和宿萼果进行定性判别分析,光谱范围在长波近红外区域(9091～4000 cm^-1),并采用原始光谱建立的校正模型对校正集正确判别率为100%,预测集的正确判别率为70%。因此,利用DA判别分析法所建的校正模型比使用SIMCA法在对库尔勒香梨的脱萼果和宿萼果进行判别时具有更高的预测精度。结果表明,利用近红外光谱也能对库尔勒香梨脱萼果和宿萼果进行自动判别。（3）基于图像处理技术的库尔勒香梨果梗提取方法研究。通过对RGB颜色模型的R、G、B颜色分量灰度进行分析,采用R-B颜色因子对背景进行分割,在香梨图像经过背景分割和边缘检测获取边缘图像后,提出一种利用数学形态学方法对香梨果梗自动提取的新方法。实验研究结果表明,该方法对香梨果梗提取的正确率为90.7%。结果表明,利用图像处理技术能对库尔勒香梨果梗实现自动提取。（4）基于近红外光谱分析技术的库尔勒香梨糖度的定量和定性分析研究。对比分析了应用偏最小二乘回归和主成分回归结合不同波段、不同预处理方法对库尔勒香梨糖度定量分析模型性能的影响。偏最小二乘回归模型的性能都明显高于主成分回归模型的性能。研究结果表明当采用偏最小二乘回归模型时,光谱范围在(12400～4000 cm^-1),主成分因子数为7,并采用MSC对原始光谱进行校正时建立的校正模型较优,所建校正模型的相关系数r_cal,为0.964,校正集的RMSEC为0.518,预测均方根误差RMSEP为0.324。依据建立的校正模型根据库尔勒香梨分类标准（NY/T 585-2002）对其进行分类,研究结果表明其对全部实验样本的特级果的正确分类率为86.36%,对一级果的正确分类率为61.54%,对二级果的正确分类率为88.89%,对等外果的正确分类率为80.00%。结果表明,利用近红外光谱可对库尔勒香梨糖度进行检测。（5）基于图像处理技术的水果重量预测模型研究。通过图像处理技术获取香梨的侧面投影面积和顶部投影面积,分别利用侧面投影面积、顶部投影面积、侧面投影面积与顶部投影面积与重量建立回归方程,研究结果表明利用侧面投影面积和顶部投影面积所建立的多元线性回归方程效果最佳,其回归方程为（?）=-47.3213+2.474232x₁+3.03103x₂,模型的决定系数为0.975,利用预测集样本对其进行检测,其相对重量的平均预测误差为2.74%,而对库尔勒香梨的重量分级率为91.30%。结果表明,利用图像处理技术可对库尔勒香梨的重量进行预测。（6）基于图像处理的水果形状识别方法研究。经过图像预处理、背景分割、利用数学形态学方法获取果梗、进而利用图像减法运算获取香梨果实图像。在正确提取果实边缘的基础上,比较了利用离散指数、人工神经网络、支持向量机和模糊C-均值聚类四种不同方法对水果形状分类的结果,识别结果如下:基于离散指数对水果的形状识别准确率为88.28%;使用离散指数、圆度作为BP神经网络输入变量,隐含层为1,研究结果表明采用TRAINRP训练函数对BP神经网络对训练集样本的识别准确率为90.91%,预测集样本识别准确率为89.66%;利用最小二乘支持向量机（LS-SVM）对水果的形状进行分类识别,研究结果表明对训练集样本的识别准确率为91.92%,预测集样本识别准确率为89.66%;利用无监督的模糊C均值聚类对水果的总的识别率为86.72%。从对库尔勒香梨形状识别方法研究结果得出,使用基于最小二乘支持向量机（LS-SVM）的水果形状识别模型的识别效果最优。结果表明,利用图像处理技术可对库尔勒香梨的形状进行识别。（7）构建适于水果品质检测的高光谱图像分析系统（400～1000 nm）,并开展了基于高光谱图像的库尔勒香梨糖度检测方法研究。在光谱范围422～982 nm范围内分析不同预处理方法对偏最小二乘回归所建模型性能的影响。与原始光谱所建模型相比,光谱经一阶微分、标准归一化（SNV）、Norries滤波等运算后所建模型相关系数有了明显提高;而模型经过二阶微分+Norris滤波和标准归一化预处理后校正模型的相关系数从0.803提高到0.898,而RMSEP则从0.644下降到0.596;模型的RMSECV从0.720下降到0.704。说明该模型比原始光谱所建立模型要好。综合比较最后选择经过标准归一化+过二阶微分和Norris滤波处理后建立的校正模型最优。结果表明,利用高光谱成像技术可对库尔勒香梨的糖度进行预测。更多还原

【Abstract】 Since current methods for detecting and grading Kuerle fragrant pear are behindhand, inefficient and destructive for its internal quality,this study aimed to develop nondestructive methods to determine the external and internal quality（sorts,weight,sugar content,fruit stalks, shape and so on） of Kuerle fragrant pear,comprehensively utilizing the knowledge of image processing technologies,spectroscopy analysis technologies,hyperspectral imaging technologies,pattern recognition,chemometrics,optics and physiology of fruit trees and so on. On the basis of the above study,rapid detection methods and models for fruit quality were established,which will provide theoretical support for developing devices of online nondestructive detection of Kuerle fragrant pear.Moreover,the research findings in this study can be used for other similar fruits.Main contents and conclusions of this study were listed as follows:（1） The discriminant analysis of the detaching calyx fruit and persistent calyx fruit were carried out by using image processing technologies.Characteristics of calyx that located at the bottom of pears were extracted by using image processing technologies,and their roundness was calculated.Thereafter,the discriminant analysis of detaching calyx fruits and persistent calyx fruits was carried out according to the threshold which was determined by statistical properties.The results indicated that the recognition accuracy for detaching calyx fruit and persistent calyx fruit were 79.7%and 85.9%.The recognition accuracy for all samples of fragrant pears was 82.8%.It can be concluded that the detaching calyx fruit and persistent calyx fruit can be classified based on image processing technologies.（2） The discriminant analysis of detaching calyx fruit and persistent calyx fruit was carried out based on NIR spectroscopy.The performances of models for qualitative analysis were compared by applying discriminant analysis（DA） and SIMCA methods with different bands and different preprocessing methods.The result indicated that the DA model,using NIR analysis technique（800-2500 nm） and DA method combined with the scope of the original spectrum(9091～4000 cm^-1) was optimal.The classification accuracy of the calibration set was 100%,and the accuracy of prediction set was 95%.When the SIMCA method and NIR-diffuse reflectance spectra(9091～4000 cm^-1) were adopted in the qualitative analysis of detaching calyx fruit and persistent calyx fruit,the accuracy of classification of correction set was 100%for the model that established with the original spectrum,and the accuracy of classification of prediction set was 70%.Accordingly,in the discriminant of detaching calyx fruit and persistent calyx fruit of Kuerle fragrant pear,the calibration model built by DA method had higher prediction accuracy than that was built by SIMCA method.The results indicate that the detaching calyx fruit and persistent calyx fruit can be classified based on NIR spectroscopy technologies.（3） The extraction methods for fruit stalk images of Kuerle fragrant pear was studied by using imaging processing technologies.RGB color models were used for analysis of the gray scale distribution of RGB component and the background were segmented by R-B color factor.After background segmentation of fragrant pears’ images and edge detection,The edge images of the fragrant pear were acquired after background segmentation and edge detection,and then a new method based on mathematical morphology was proposed for automatic extraction of stalks of fragrant pears.The experimental results showed that the accuracy rate of this method for extracting stalks of fragrant pears was 90.7%.It can be concluded that the fruit stalk of Kuerle fragrant pear can be extracted automaticly based on image processing technologies.（4） The quantitative and qualitative analysis for sugar content of Kuerle fragrant pears was carried out based on NIR spectroscopy.The performances of models developed by partial least squares regression（PLSR） and principal component regression（PCR） for sugar content were compared coupled with different bands and different preprocessing methods. The performance of PLSR was obviously better than that of PCR.The results of the study showed that when PLSR model was applied,in the range of 12400～4000 cm^-1,with seven principal component factors,the optimal correction model could be generated with the original spectra corrected by MSC.The correlation coefficient of the model,r_cal,was 0.964 and the RMSEC of calibration set was 0.518.The RMSEP was 0.324.The correction model established with standard of classification for Kuerle fragrant pear（NY/T 585-2002 ） was used.The results showed that overall correct classification rate of high-grade,grade 1, grade 2 and low-grade were 86.36%,61.54%,88.89%,and 80%,respectively.The results indicate that the sugar content of Kuerle fragrant pear can be determinated based on NIR spectroscopy technologies.（5） The prediction model for weight of fruits,which were based on image processing,was studied.The area of side projection and top projection were acquired using image processing technologies.The regression equation,（?）=-47.3213+2.474232x₁+3.03103x₂, was established on the basis of relationship between side & top projection area and the weight,and the coefficient of determination was 0.975.The fruits were determined using prediction samples.The average prediction error of weights was 2.74,and the rate of weight classification for Kuerle fragrant pear was 91.30%.It can be concluded that the weight of Kuerle fragrant pear can be predicted based on image processing technologies.（6） The study on recognition methods of fruit shape based on digital image processing was described.After the process of image preprocessing and background segmentation,images of fruit stalks were obtained by morphology strategies,and therefore images of fragrant pear were acquired through application of image subtraction.On the basis of proper extraction of fruit edges,four different methods,using discrete index,artificial neural networks（ANN）,support vector machines（SVM）,and fuzzy C-means（FCM） clustering algorithm,were compared for classification of fruit shapes.Identification results were as follows:the accuracy of fruits’ shape recognition based on discrete index was 88.28%. When discrete index and roundness were used as input variables in BP neural network and the number of hidden layer nodes was 1,results showed the accuracy recognition for training set was 90.91%when TRAINRP train function was used in BP network,and the accuracy recognition for prediction set was 89.66%.When LS-SVM was used for classification and recognition of fruit shapes,results showed that the accuracy recognition for training set was 91.92%,and the accuracy recognition for prediction set was 89.66%.In the case of unsupervised FCM clustering,the total recognition rate of fruits was 86.72%. Results of this study on recognition methods of Kuerle fragrant pear showed that performance of model which was based on LS-SVM was optimal among the recognition methods of fruits’ shape.The results indicate that the shape of Kuerle fragrant pear can be identified based on image processing technologies.（7） An experimental platform of hyperspectral imaging for detection of fruit quality was established（400-1000 nm）,and the research method for testing sugar content of Kuerle fragrant pear was developed.In the spectral range of 422～982 nm,the influence of different preprocessing methods on performance of the model,which was established by PLSR,were evaluated.By comparison,the relation coefficient of models,which were treated by first-order differential,standard normal variate（SNV）,Norries filter and other algorithms,were much higher than that of model of original spectrum.When models were preprocessed by second-order differential,Norris filter and SNV,the coefficient of relation was increased from 0.803 to 0.898,and the RMSEP was decreased from 0.644 to 0.596. The RMSECV of model was decreased from 0.720 to 0.704.It indicated that this model was better than that established with original spectrum.After comprehensive comparison, the corrected model,which was treated with normalization,second-order differential and Norris filter,was optimal.The result of this study showed that soluble solid content of Kuerle fragrant pear could be predicted by hyperspectral imaging technologies.It can be concluded the sugar eontent of Kuerle fragrant pear can be predicted based on hyperspectral imaging technologies.更多还原