【作者】 傅霞萍；

【导师】 应义斌；

【作者基本信息】 浙江大学 ， 生物系统工程， 2008， 博士

【摘要】 水果是人类饮食结构的基本组成部分,为人体提供了丰富的营养物质。我国是水果生产大国,但并不是水果生产强国,产后加工和处理水平低是导致我国水果品质差、国际市场竞争力弱的主要原因之一,因而实现水果外观和内部品质的快速无损检测及分级已经成为我国水果产业化的必要前提。本研究是对本课题组前期研究(基于机器视觉技术的水果外观品质检测和基于近红外(near infrared, NIR)光谱技术的水果糖、酸度检测)的进一步深入和拓展,目的是为实现水果品质的多指标综合评价。本课题以梨和猕猴桃为研究对象,利用可见／近红外光谱分析技术、光纤传感技术和化学计量学分析方法,结合理化分析、水果生理和病理知识,开展水果内部品质主要包括坚实度、维生素C含量的检测研究,并在此基础上建立各品质指标的近红外光谱定量预测模型。本文还利用可见／近红外光谱分析技术结合模式识别方法,对水果内部缺陷、水果品种进行鉴别研究,并建立相应的定性判别或分类模型。本研究的主要内容、结果和结论为：(1)探索了光谱采集参数对分析结果的影响。解释了Nexus智能型傅立叶变换(Fourier transform, FT)光谱仪在光谱采集时增益、动镜速度和光圈大小等参数的设定依据,并分析了扫描次数和分辨率对光谱和建模结果的影响,结果显示：模型性能随分辨率增加而提高,但扫描时间也随之增加,而扫描次数为64次的时候模型的性能相对较好；分析了用USB4000便携式微型光纤光谱仪采集光谱时积分时间和平均次数等参数的设置依据；(2)分析了原始光谱及经预处理后的光谱(包括微分光谱和平滑光谱)的谱线特征,结果显示水果的漫反射原始光谱特征吸收主要在970 nm.1190 nm.1450 nm.1790 nm和1940 nm附近,与O-H和C-H基团的振动和倍频吸收相关；水果的原始透射光谱在680-690 nm.790-800 nm附近的透过率变化比较显著,具有内部褐变缺陷的梨在柄-蒂垂直放置采集的光谱低于750 nm时比完好的梨具有更强的吸收,而在高于750 nm时吸收变弱；在柄-蒂水平放置采集的光谱的特性与柄-蒂垂直放置时的光谱的特性基本一致,只是这一分界点移到了720 nm左右。对平滑光谱的平均吸光度和均方根噪声的方差分析结果显示：不同点数平滑处理对两台光谱仪器所采集的光谱的特征信息的影响不显著。(3)分析了水果不同检测部位的光谱信息的差异,对同一水果三个纬度和三个经度的光谱平均吸光度和均方根噪声的方差分析结果显示：水果不同纬度上的光谱差异较大而不同经度上的差异较小。(4)基于Chauvenet检验方法对各组样本中的光谱异常样本进行分析,在剔除光谱异常样本后进行梨内部缺陷判别和梨品种分类的定性建模分析；基于杠杆值和学生残差T检验方法对各组样本中的浓度异常样本进行分析,将杠杆值和学生残差值较大的样本暂定为异常样本,通过比较逐一回收后的模型性能确定最终剔除的浓度异常样本,定量分析所用样本同时基于光谱异样本剔除结果和浓度异常样本剔除结果。(5)研究了梨内部缺陷的定性判别分析：比较不同仪器和检测方式的判别结果显示,用USB4000微型光纤光谱仪采集的透射光谱建立的模型的判别正确率要高于用FT-NIR光谱仪采集的漫反射光谱建立的模型,由此可以认为用透射光谱进行梨内部缺陷判别的效果要优于漫反射光谱；比较判别分析(Discriminant analysis, DA)、簇类的独立软模式分类法(Soft independent modeling of class analogy, SIMCA)、判别偏最小二乘法(Discriminant partial least squares, DPLS)及概率神经网络(Probobilistic neural network, PNN)四种判别方法对梨内部缺陷的判别结果显示,DPLS模型和PNN模型不适合用于梨内部缺陷的判别分析,SIMCA模型的判别结果略优于DA模型；对两种水果放置方式(柄-蒂垂直和柄-蒂水平)的比较结果显示,在水果柄-蒂水平放置时所采集的透射光谱对缺陷的判别效果更优；比较多种光谱预处理方法和不同建模波段对梨内部缺陷判别结果的影响显示,对储藏期的雪青梨,较优的模型是基于水果水平放置时获得的450-1000 nm经平滑处理后的透射光谱所建的SIMCA模型,校正集和预测集的判别正确率分别为92.68%和78.57%；对储藏期的翠冠梨,较优的模型是基于水果水平放置时获得的450-1000 nm经多元散射校正(Multiplicative scatter correction, MSC)处理后的透射光谱所建的SIMCA模型,校正集和预测集的判别正确率分别为96.15%和88.24%。(6)研究了不同品种梨的定性分类分析：比较不同仪器和检测方式的分类结果显示,漫反射光谱的分类效果要优于透射光谱,由USB4000光纤光谱仪获得的漫反射光谱的校正模型的分类效果要优于由傅立叶变换光谱仪获得的漫反射光谱的校正模型,但其预测效果不如InGaAs检测器所获得的漫反射光谱模型的预测效果,综合考虑校正集和预测集的分类正确率,用InGaAs检测器获得的800-2500 nm的近红外漫反射光谱的分类效果最优；比较DA、SIMCA、DPLS及PNN四种方法对不同品种梨的分类结果显示,DA和SIMCA两者的分类结果比较接近的,且明显优于DPLS模型分类结果,PNN校正模型的分类效果与DA或SIMCA分类模型的分类效果相近,但PNN预测模型的预测性能不及DA和SIMCA模型的预测性能,综合考虑校正集和预测集的分类正确率,DA模型的性能最优；比较多种光谱预处理方法和不同建模波段对不同品种梨分类结果的影响显示,基于1100-2500 nm的漫反射光谱经25点平滑后所建立的DA综合模型的分类效果最好,校正集和预测集的分类正确率分别为99.43%和99%。(7)研究了梨坚实度的定量分析：比较偏最小二乘回归(Partial least squares regression, PLSR)、主成分回归(Principal components regression, PCR)、多元线性回归(Multi linear regression, MLR)及最小二乘支持向量机(Least square support vector machine, LS-SVM)四种定量校正方法对梨坚实度的检测结果显示,PCR模型和MLR模型的性能总体上不如PLSR模型；基于PCA的LS-SVM模型在包含10个主成分时的模型性能与PSLR原始光谱全波段模型性能相近,05年西子绿梨的LS-SVM模型的校正集和预测集相关系数分别为0.870、0.849, SEC和SEP分别为2.85 N、2.78 N,05年翠冠梨的LS-SVM模型的校正集和预测集相关系数分别为0.943.0.731, SEC和SEP分别为1.15 N、1.98 N,05年雪青梨的LS-SVM模型的校正集和预测集相关系数分别为0.898、0.774, SEC和SEP分别为1.78 N、2.41N;比较多种光谱预处理方法和不同建模波段对梨坚实度检测结果的影响显示,05年西子绿梨的光谱经15点平滑处理后在800-1880 nm和2190-2220 nm所建PLSR模型的性能较优,校正和交互验证的相关系数分别为0.916和0.746, SEC、SEP和SECV分别为2.25 N、3.26 N和3.77 N；05年翠冠梨的光谱经SNV校正处理后在1374-1565 nm、1814-1894 nm和2017-2217 nm所建PLSR模型的性能较优,校正和交互验证的相关系数分别为0.922和0.731, SEC、SEP分别为SECV和1.22 N、2.02 N年雪青梨的光谱2.18 N; 05经15点平滑处理后在所建800-2500 nm模型的性能较优,校正和交互验证的PLSR相关系数分别为0.893和和0.808, SEC、SEP分别为SECV和2.31 N。1.75 N、2.32 N(8)研究了梨坚实度的定量模型修正：用斜率／截距法对05年采收期梨坚实度模型的修正结果显示,对于无预处理模型和经平滑、微分和散射校正处理后的模型,经修正后的模型预测结果均有明显改善,预测误差SEP比未修正之前明显减小,修正后坚实度实际值与预测值的差值分布比修正前更接近于零线。(9)研究了猕猴桃维生素C含量的定量分析：比较PLSR、PCR、MLR及LS-SVM四种定量校正方法以及多种光谱预处理方法和不同建模波段对猕猴桃维生素C含量的检测结果显示,用猕猴桃原始光谱在全波段范围建立的PLSR模型对维生素C含量的预测性能较优,校正集和预测集的相关系数分别为0.932、0.616,SEC、SEP和SECV分别为7.95 mg/100g.15.8 mg/lOOg和17.5 mg/100g; PCR模型的性能总体上不如PLSR模型；用11个波长的光谱信息建立MLR模型的校正集相关系数达到0.9以上,基本接近PLSR模型的校正性能,交互验证的性能则明显优于PLSR模型,不过预测性能比较差；基于PCA的LS-SVM模型随着所含主成分数的增加,性能不断地改善,但总体性能不如PLSR模型；用PLS因子权重替代主成分建立的LS-SVM模型的性能也随着模型所含因子数的增加而不断地改善,与PLSR较优模型使用相同因子数时,校正集和预测集的相关系数分别为0.926和0.907,SEC和SEP分别为8.44mg/100g和8.78mg/100g,虽然校正性能略微变差,但预测性能明显提高。更多还原

【Abstract】 Fruit is one of the main components of human diet. It provides abundant nutritional elements for human body. China is a big fruit producer, but not powerful. One of the important reasons is the low fruit quality and weak competitiveness in world market caused by low commercialization treatment ability for post-harvest fruit. Therefore, rapid and nondestructive detection and classification of fruit external and internal qualities becomes necessary for fruit industrialization of our country. This study is a further extension of the former research projects done by our group members-fruit external quality detection using machine vision technique and fruit sugar content and acidity detection using near infrared (NIR) spectroscopy, in order to evaluate the fruit quality with multiple quality indices.The research objects are pears and kiwifruits. Detection of pear firmness and kiwifruit vitamin C content were studied using visible/NIR spectroscopy, optic fiber sensor and chemometrics techniques combined with physical-chemical analysis and fruit physiological and pathological knowledge. Quantitative models were established based on visible/NIR spectra for fruit firmness and vitamin C content determination. In this dissertation, pear internal defect discriminant and pear variety classification were also studied using visible/NIR spectroscopy technique and pattern recognition methods. Qualitative models were established for pear internal defect discriminant and variety classification.The main results and conclusions were:1. The influence of spectra acquisition parameters on spectra and modeling results were analyzed. For FT-NIR spectrometer, the setting principal of gain, moving mirror velocity and aperture parameters were explained. The influence of scan number and resolution on spectra and modeling results were analyzed. The results indicated that: model performance was improved with resolution increasing, but the time cost for scanning also increased. Models were relatively better when the scan number was 64. For USB2000/4000 miniature optic fiber spectrometer, the setting principal of integral time and average time were also explained. 2. The curve characteristics of fruit original spectra and spectra after derivative and smoothing pretreatments were analyzed. Main absorption properties on fruit diffuse reflectance spectra were around 970 nm,1190 nm,450 nm,1790 nm and 1940 nm, which were related to O-H and C-H functional groups. The transmittance changes on transmission spectra were much obvious around 680-690 nm and 790-800 nm. With fruit stem-calyx vertical, spectra of pears with internal defect have absorbed more light than sound pears below 750 nm, and absorbed less light than sound pears above 750 nm. With fruit stem-calyx horizontal, the characteristics of defect pear and sound pear were nearly the same, but the threshold moved to 720 nm. The results of variation analysis (ANOVA) of average absorbance and root mean square noise for original spectra and smoothed spectra showed that the spectra characteristics were not influenced by smoothing pretreatments with different points for both FT-NIR and USB2000/4000 spectrometers.3. The difference of spectra acquired from different fruit locations was analyzed. The results of ANOVA of average absorbance and root mean square noise for spectra acquired from three latitudes and three longitudes (nine points) on each fruit shown that the difference of spectra from different latitudes were much more than the difference of spectra from different longitudes.4. Spectra outliers in each sample set were analyzed by Chauvenet testing method. Qualitative models for pear internal defect discriminant and pear variety classification were established after eliminating spectra outliers. Concentration outliers in each sample set were analyzed by leverage and student residual testing method. Samples with relatively larger leverage value or student residual value were considered as outlier and removed from the sample set firstly. And then they were reclaimed to the model one by one to see whether they provided any useful information or not. Quantitative models for pear firmness and kiwifruit vitamin C content detection were established based on both spectra outlier elimination and concentration outlier elimination. 5. Qualitative analysis of pear internal defect: Comparison results of different spectrometers (or detectors) and detection modes indicated that discriminant accuracy of models based on transmission spectra using USB4000 miniature optic fiber spectrometer were much better than the accuracy of models based on diffuse reflectance spectra using FT-NIR spectrometer, which can be concluded that transmission spectra were more suitable for pear internal defect discriminant. Comparison results of different pattern recognition methods of discriminant analysis (DA), soft independent modeling of class analogy (SIMCA), discriminant partial least squares (DPLS), and probabilistic neural networks (PNN) indicated that DPLS and PNN models were not suitable for pear internal defect discriminant, and the discriminant results of SICMA model were a bit better than those of DA model. Comparison results of two fruit placing modes (stem-calyx vertical and stem-calyx horizontal) indicated that discriminant accuracy of the model using spectra acquired with fruit stem-calyx horizontal was much better. Comparison results of different spectra pretreatments and different wavebands showed that:For Xueqing pears after storage, the SIMCA model based on spectra acquired with fruit stem-calyx horizontal in 450-1000 nm after smoothing pretreatment was much better. The discriminant accuracy of calibration and validation were 92.68%and 78.57%, respectively. For Cuiguan pears after storage, the SIMCA model based on spectra acquired with fruit stem-calyx horizontal in 450-1000 nm and using multiplicative signal correction (MSC) pretreatment was much better. The discriminant accuracy of calibration and validation were 96.15%and 88.24%, respectively.6. Qualitative analysis of pear variety: Comparison results of different spectrometers (or detectors) and detection modes indicated that classification correctness of models based on diffuse reflectance spectra were much better than that of models based on transmission spectra. According to the correctness of both calibration and validation, the best model was based on diffuse reflectance spectra acquired by InGaAs detector in the range of 800-2500 nm. Comparison results of models using DA, SIMCA, DPLS, and PNN methods indicated that the classification correctness of DA and SIMCA models were very close, and were much better than that of DPLS model. The calibration results of PNN model were close to those of DA and SIMCA models, but the prediction results were much worse. According to the correctness of both calibration and validation, the performance of DA model was the best. Comparison results of different spectra pretreatments and different wavebands showed that:DA model using diffuse reflectance spectra in 1100-2500 nm and after 25 points smoothing pretreatments turned out the best results, with classification correctness of 99.43%and 99%for calibration and validation, respectively.7. Quantitative analysis of pear firmness: Comparison results of different calibration methods of partial least square regression (PLSR), principal components regression (PCR), multi linear regression (MLR), and least square support vector machines (LS-SVM) showed that the performance of PLSR model was better than PCR model and MLR model. The performance of LS-SVM model using ten principal components was close to that of PLSR model. For Xizilv pears (2005), the correlation coefficients of calibration and validation were 0.870 and 0.849, respectively; standard error of calibration (SEC) and standard error of prediction (SEP) were 2.85 N and 2.78 N, respectively. For Cuiguan pears (2005), the correlation coefficients of calibration and validation were 0.943 and 0.731, respectively; SEC and SEP were 1.15 N and 1.98 N, respectively. For Xueqing pears (2005), the correlation coefficients of calibration and validation were 0.898 and 0.774, respectively; SEC and SEP were 1.78 N and 2.41 N, respectively.Comparison results of different spectra pretreatments and different wavebands showed that:For Xizilv pears (2005), the performance of PLSR model established in 800-1880 nm and 2190-2220 nm after 15 points smoothing pretreatments was much better, the correlation coefficients of calibration and validation were 0.916 and 0.746, respectively; SEC, SEP and standard error of cross validation (SECV) were 2.25 N, 3.26 N and 3.77 N, respectively. For Cuiguan pears (2005), the performance of PLSR model established in 1374-1565 nm,1814-1894 nm and 2017-2217 nm after standard normal variate (SNV) correction was much better, the correlation coefficients of calibration and validation were 0.922 and 0.731, respectively; SEC, SEP and SECV were 1.22 N,2.02 N and 2.18 N, respectively. For Xueqing pears (2005), the performance of PLSR model established in 800-2500 nm after 15 points smoothing pretreatments was much better, the correlation coefficients of calibration and validation were 0.893 and 0.808, respectively; SEC, SEP and SECV were 1.75 N,2.32 N and 2.31 N, respectively.8. Quantitative model modification for pear firmness prediction: The modification results of pear firmness models using slope/incept method indicated that the predictive performance of models after modification was improved obviously compared to that of models before modification, not only for spectra without pretreatments but also for spectra with smoothing, derivative, or scattering correction pretreatments. After modification, SEP decreased distinctively, and the difference distribution of actual and predicted firmness become closer to zero line.9. Quantitative analysis of kiwifruit vitamin C content: Comparison results of different calibration methods of PLSR, PCR, MLR and different spectra pretreatments and different wavebands showed that the performance of PLSR model using original spectra in 800-2500 nm was much better, the correlation coefficients of calibration and validation were 0.932 and 0.616, respectively; SEC, SEP and SECV were 7.95 mg/100g,15.8 mg/100g and 17.5 mg/100g, respectively. PCR model performed worse than PLSR model. The correlation coefficient of MLR calibration using eleven wavelengths was above 0.9, which is close to PLSR calibration performance, and the cross validation performance was better than PSLR model, however, the prediction performance was much worse. The performance of LS-SVM model based on principal component analysis (PCA) was improved with more principal components included, but it was also worse than PLSR model.Using PLS factor loadings instead of principal component scores, the performance of LS-SVM model can be improved and also increased with more factors included. When LS-SVM using the same factors as PLSR model, the correlation coefficients of calibration and validation were 0.926 and 0.907, respectively; SEC and SEP were 8.44 mg/100g and 8.78 mg/100g, respectively. Although, the calibration performance was a bit worse than PLSR model, the prediction performance was improved distinctively.更多还原