基于光谱和图像信息融合的哈密瓜成熟度无损检测研究

发布时间：2018-06-27 15:46

本文选题：哈密瓜 + 成熟度　；参考：《石河子大学》2017年博士论文

【摘要】：哈密瓜是新疆特产厚皮甜瓜品种群的统称,其果实甘甜香脆、营养丰富、并且具有食疗作用,备受广大消费者青睐,有“中华第一蜜瓜”的美称。随着人们生活水平的不断提高,消费者逐渐对哈密瓜内部品质和外观品质越来越重视。目前对哈密瓜内部品质检测还采用传统的破坏性检测方法,其效率低、耗时长;对哈密瓜成熟度判别主要依靠人为主观经验判断,其判定结果随意性大,误差大,缺乏可靠性和准确性,已不适于大规模的生产需要。针对这些问题,本文以不同成熟度的哈密瓜为研究对象,利用近红外光谱技术、高光谱成像技术、计算机视觉技术对哈密瓜内部品质和成熟度同时进行无损检测研究。研究了哈密瓜近红外光谱和计算机视觉图像特征的融合方法,并最终建立了哈密瓜成熟度判别融合模型。主要研究内容如下:1.分析了哈密瓜内部品质(可溶性固形物、总酸、硬度)与成熟度的相关性。哈密瓜的成熟度是果实发育过程中各个理化指标变化的一个综合性的评判结果。为了寻找最能反映哈密瓜成熟度的表征因子,对哈密瓜的可溶性固形物(SSC)、总酸(TAC)和硬度与其成熟度作了Pearson相关性分析和单因素方差分析。结果表明,SSC、TAC和硬度与成熟度之间存在显著相关(p0.01),其SSC和成熟度的相关性最高,且不同成熟度哈密瓜之间的SSC差异最大,因此为选择合适的哈密瓜成熟度表征因子提供了理论依据。2.建立了基于近红外光谱技术的哈密瓜内部品质预测模型。研究比较了四种光谱预处理方法(Norm、SNV、MSC、SG-1-Der)对哈密瓜SSC、TAC及硬度的PLS和非线性SVR模型预测结果。研究表明,Norm和MSC预处理方法的预测结果较好。为了筛选重要的特征波长来简化模型和提高模型预测能力,利用四种特征波长筛选方法(UVE、GA、SPA、CARS)筛选了各指标的特征波长,并建立了PLS和非线性SVR预测模型。研究表明,PLS模型预测结果优于SVR模型,GA方法能够筛选出有效的哈密瓜SSC、TAC和硬度的特征波长,其各指标的GA-PLS模型预测结果均最优。SSC-GA-PLS模型预测集相关系数(Rp)、均方根误差(RMSEP)和RPD分别为0.9061、0.4630和2.363,Rp较原始光谱提高了4.66%;TAC-GA-PLS模型Rp、RMSEP和RPD分别为0.8277、0.0391和1.781,其Rp较原始光谱模型提高了3.44%;硬度的GA-PLS模型Rp、RMSEP和RPD分别为0.8729、28.47和2.049,Rp较原始光谱模型提高了1.2%。3.建立了基于近红外光谱技术的哈密瓜成熟度判别模型。研究比较了不同光谱预处理方法和不同建模方法的哈密瓜成熟度判别结果。研究表明,MSC和SNV方法处理后的判别结果较优,非线性SVM模型对哈密瓜成熟度的判别结果略优于线性的LDA、PLS-DA和SIMCA模型,其中SNV-SVM模型校正集和预测集判别结果分别为94%和88%。对比分析了利用哈密瓜SSC和成熟度的特征波长结合上述四种建模方法对哈密瓜成熟度的判别结果,结果表明,基于成熟度特征波长的判别结果优于哈密瓜SSC特征波长的成熟度判别结果。利用SPA方法从全波段中筛选的7个哈密瓜成熟度特征波长(914.75 nm,921.26 nm,934.28 nm,940.79 nm,1318 nm,2132.9 nm,2364.8 nm)所建立的SVM模型最优,其校正集和预测集判别结果分别为93%和86%,模型仅用了7个特征波长变量,占全光谱信息的2.73%,其校正集和预测集判别结果较全光谱SNV-SVM模型仅降低了1%和2%。4.建立了基于高光谱成像技术的哈密瓜内部品质预测模型。对比分析了不同光谱预处理方法对哈密瓜SSC、TAC及硬度的PLS和非线性SVR预测模型的影响。结果表明,SNV和MSC预处理方法性能较优。在哈密瓜SSC预测模型中,SNV-SVR模型预测结果略低于MSC-PLS模型,而TAC和硬度的SVR模型的预测结果较PLS模型分别提高了约3%和9%。比较了联合区间偏最小二乘法(si-PLS)、GA、SPA和CARS四种光谱特征波长筛选方法对PLS和非线性SVM预测模型精度的影响。研究表明,CARS方法能够有效地筛选哈密瓜SSC、TAC和硬度特征波长,筛选的特征波长变量所建立各指标模型的预测结果均较原始光谱有所提高。结果表明,非线性SVR模型预测结果略优于PLS模型,SSC-CARS-SVR模型预测结果最优,其Rp、RMSEP和RPD分别为0.9394、0.4071和2.917,Rp较原始光谱模型提高了2.36%;TAC-CARS-SVR模型预测结果最优,其Rp、RMSEP和RPD分别为0.8705、0.0322和2.032,其Rp较原始光谱模型提高了7.79%;在哈密瓜硬度预测模型中,四种方法所筛选的特征波长的SVM模型预测结果均低于全光谱MSC-SVM模型。5.建立了基于高光谱成像技术的哈密瓜成熟度判别模型。对比分析了不同光谱预处理方法和不同建模方法的哈密瓜成熟度判别结果。结果表明,SNV预处理方法性能最优,非线性SVM模型对哈密瓜成熟度的判别结果优于线性的LDA、PLS-DA和SIMCA模型,其中SNV-SVM模型最优,其校正集和预测集判别结果分别为98%和94%。比较了利用哈密瓜SSC和成熟度的特征波长结合四种建模方法对哈密瓜成熟度判别结果。研究表明,基于成熟度特征波长的判别结果优于哈密瓜SSC特征波长的判别结果,利用CARS方法筛选的57个哈密瓜成熟度特征波长所建立的SVM模型最优,其校正集和预测集判别结果分别为100%和95%。结果表明,CARS方法能够较好的筛选到反映哈密瓜成熟度的关键特征波长变量,不但简化了模型,而且提高了模型预测能力。6.在高光谱数据中的光谱特征波长筛选研究中,采用SPA方法在CARS筛选的57个特征波长中再次优选了8个哈密瓜成熟度特征波长(552.21 nm、585.19 nm、608.32 nm、644.74 nm、734.16 nm、815.11 nm、859.85 nm和993.2 nm),用灰度共生矩阵法提取了特征波长图像的纹理特征,分别建立了基于哈密瓜特征波长、纹理特征、特征波长和纹理特征融合的哈密瓜成熟度SVM判别模型。结果表明,光谱模型判别结果优于图像特征模型,光谱和图像特征融合的SVM模型最优,其校正集判别结果为100%,预测集判别结果高达97%。7.建立了基于计算机视觉技术的哈密瓜成熟度的判别模型。提取了哈密瓜图像在RGB和HSV颜色空间中的6个颜色分量,并研究了HSV颜色空间中色调H与哈密瓜成熟度的关系。利用灰度共生矩阵法提取了6个哈密瓜图像纹理特征,结合二值化纹理图像计算了哈密瓜纹理覆盖率。对比分析了哈密瓜图像颜色特征、纹理特征、颜色与纹理特征融和的哈密瓜成熟度的LDA、PLS-DA、SIMCA和SVM模型判别结果。结果表明,SVM建模方法最优,单一的颜色特征模型判别结果优于纹理特征模型;基于颜色特征和纹理特征融合的SVM模型判别结果最优,其校正集和预测集判别正确率分别为96%和92%。8.建立了基于光谱及图像信息融合的哈密瓜成熟度判别模型。比较了DS证据理论、ELM极限学习机、SVM支持向量机和Ada Boost分类器4种融合方法的哈密瓜成熟度建模结果。基于决策层融合的DS证据理论建立的哈密瓜成熟度判别模型的校正集和预测集判别结果分别为96%和92%,与单一的计算机视觉颜色特征模型最优判别结果相同。在三个特征层融合模型中,其校正集判别正确率均为100%,SVM支持向量机和Ada Boost分类器建立的融合模型的预测集判别结果均为97%,ELM极限学习机融合模型最优,其预测集判别结果为98%。结果表明,ELM极限学习机特征层融合为最佳融合方法,建立的模型为最优融合模型。9.对比分析了信息融合技术与单一的近红外光谱技术、计算机视觉技术和高光谱成像技术所建立的哈密瓜成熟度判别结果。结果表明,高光谱成像技术所建的哈密瓜成熟度判别模型的判别结果优于近红外光谱和计算机视觉技术判别模型;计算机视觉技术所建模型判别结果优于近红外光谱技术;近红外光谱特征与计算机视觉图像融合所建的模型最优,预测集判别正确率为98%,较高光谱、近红外光谱和计算机视觉技术模型分别提高了1%,12%和8%。研究表明,利用光谱及图像信息融合技术判别哈密瓜成熟度是可行的,本文的研究结果为同时对哈密瓜内部品质预测及成熟度进行快速无损检测提供一些思路和方法,具有一定的参考价值。
[Abstract]:Hami melon is the general name of the cantaloupe variety of Xinjiang special thick skinned melon. Its fruit is sweet and sweet, rich in nutrition, and has a therapeutic effect. It is favored by the consumers and has the name of "the first honeydew melon". With the continuous improvement of people's living standard, consumers gradually pay more and more attention to the internal quality and appearance quality of Hami melon. The internal quality detection of Hami melon also adopts the traditional destructive testing method, which has low efficiency and long time consuming. The judgement of Hami melon maturity depends mainly on human subjective experience, and its decision results are random, large error, lack of reliability and accuracy, and are not suitable for large-scale production needs. By using near infrared spectroscopy, hyperspectral imaging technology and computer vision technology, the internal quality and maturity of Hami melon were detected at the same time. The fusion method of Hami melon near infrared spectrum and computer visual image features was studied. Finally, the fusion model of Hami melon maturity was established. The main contents are as follows: 1. the correlation between the internal quality of Hami Melon (soluble solids, total acid, hardness) and maturity is analyzed. The maturity of Hami melon is a comprehensive evaluation result of the changes in physical and chemical indexes during the development of fruit. In order to find the most capable of reflecting the maturity of Hami melon, Hami melon can be found. Soluble solid (SSC), total acid (TAC) and its hardness and its maturity were analyzed by Pearson correlation analysis and single factor variance analysis. The results showed that there was a significant correlation between SSC, TAC and hardness and maturity (P0.01). The correlation between SSC and maturity was the highest, and the difference of SSC between different mature degrees of Hami melon was the greatest. The maturity characterization factor of melons provides a theoretical basis for the establishment of the internal quality prediction model of Hami melon based on near infrared spectroscopy based on.2.. The results of four spectral preprocessing methods (Norm, SNV, MSC, SG-1-Der) for the SSC, TAC, and hardness PLS and nonlinear SVR model prediction results of Hami melon are compared. The study shows that Norm and MSC pretreatment methods The prediction results are better. In order to screen important characteristic wavelengths to simplify the model and improve the model prediction ability, four characteristic wavelength screening methods (UVE, GA, SPA, CARS) are used to select the characteristic wavelengths of each index, and the PLS and nonlinear SVR prediction models are established. The study shows that the prediction results of the PLS model are better than the SVR model, and the GA method can be screened out. Effective Hami melon SSC, TAC and hardness characteristic wavelengths, the GA-PLS model prediction results of each index are all the optimal.SSC-GA-PLS model predictive set correlation coefficient (Rp), the root mean square error (RMSEP) and RPD are 0.9061,0.4630 and 2.363 respectively, Rp is 4.66% higher than the original spectrum, TAC-GA-PLS model Rp, RMSEP and 1.781, respectively. Compared with the original spectral model, the GA-PLS model Rp, RMSEP and RPD are 0.8729,28.47 and 2.049 respectively, and Rp is higher than the original spectral model by 1.2%.3., and the maturity discrimination model of the Hami melon based on the near infrared spectroscopy is established. The study compares the maturity of the Hami melon with different spectral preprocessing methods and different modeling methods. The results of the study show that the discriminant results of the MSC and SNV methods are better. The nonlinear SVM model is slightly better than the linear LDA, PLS-DA and SIMCA model, in which the SNV-SVM model correction set and the prediction set are 94% and 88%., respectively, to analyze the characteristic wavelengths of the Hami melon SSC and maturity. Combining the four modeling methods to distinguish the maturity of Hami melon, the results show that the discriminant results based on maturity characteristic wavelengths are superior to the maturity discrimination results of SSC characteristic wavelengths of Hami melon. 7 Hami melon mature characteristic wavelengths (914.75 nm, 921.26 nm, 934.28 nm, 940.79 nm, 1318 nm) are selected from the whole band. The SVM model of 2132.9 nm, 2364.8 nm) is optimal, and the results of the correction set and the prediction set are 93% and 86% respectively. The model only uses 7 characteristic wavelength variables, accounting for 2.73% of the total spectral information. The correction set and the prediction set result are only 1% and 2%.4., which are only 1% and 2%.4., and the hyperspectral imaging technology based Hami melon is established. The influence of different spectral preprocessing methods on the SSC, TAC and hardness PLS and nonlinear SVR prediction models of Hami melon was compared and analyzed. The results showed that the performance of SNV and MSC preprocessing methods was better. In the Hami melon SSC prediction model, the prediction result of SNV-SVR model was slightly lower than the MSC-PLS model, while TAC and SVR model of the hardness. Compared with the PLS model, the prediction results were increased by about 3% and 9%. respectively. The influence of the four spectral characteristic wavelength screening methods of the joint interval partial least squares (si-PLS), GA, SPA and CARS on the precision of PLS and nonlinear SVM prediction model. The study shows that the CARS method can effectively screen the characteristic wavelengths of TAC and hardness and the characteristic wavelengths for screening of Hami melon SSC. The prediction results of each index model are better than that of the original spectrum. The results show that the prediction results of nonlinear SVR model are slightly better than that of PLS model, and the prediction results of SSC-CARS-SVR model are optimal. The Rp, RMSEP and RPD are 0.9394,0.4071 and 2.917 respectively, Rp is 2.36% higher than that of the original spectral model, and the prediction results of the TAC-CARS-SVR model are optimal. Its Rp, RMSEP and RPD are 0.8705,0.0322 and 2.032 respectively, and their Rp is 7.79% higher than the original spectral model. In the Hami melon hardness prediction model, the prediction results of the SVM model of the characteristic wavelengths selected by the four methods are lower than the full spectral MSC-SVM model.5. to establish the Hami melon maturity discriminant model based on hyperspectral imaging technology. The results of different spectral preprocessing methods and different modeling methods of Hami melon maturity discrimination results show that the SNV preprocessing method has the best performance. The nonlinear SVM model is superior to the linear LDA, PLS-DA and SIMCA model, among which the SNV-SVM model is the best, and the correction set and the prediction set are 98% respectively. Compared with the characteristic wavelength of Hami melon SSC and the characteristic wavelengths of maturity, four modeling methods are used to discriminate the maturity of Hami melon. The study shows that the discriminant result based on the characteristic wavelength of the maturity is superior to the discriminant result of the SSC characteristic wavelength of Hami melon, and the SVM model of the 57 Hami melon maturity characteristic wavelengths selected by the CARS method is used. The results of the correction set and the prediction set are 100% and 95%. respectively. The results show that the CARS method can better screen the key characteristic wavelength variables that reflect the maturity of Hami melon. It not only simplifies the model, but also improves the model prediction ability of.6. in spectral characteristic wavelength screening in hyperspectral data, and uses the SPA method in the study of the spectral characteristics of the hyperspectral data. In the 57 characteristic wavelengths selected by CARS, 8 Hami melon mature characteristic wavelengths were selected again (552.21 nm, 585.19 nm, 608.32 nm, 644.74 nm, 734.16 nm, 815.11 nm, 859.85 nm and 993.2 nm). The texture features of the characteristic wavelength images were extracted by gray-scale symbiotic matrix method, and the feature wavelengths were based on the feature wavelengths of Hami melon, texture features and characteristic waves. SVM discriminant model of Hami melon maturity with the fusion of long and texture features. The results show that the discriminant result of spectral model is superior to the image feature model, the SVM model of the fusion of spectral and image features is the best, the result of the correction set is 100%, and the prediction set is up to 97%.7. to establish the judgement of Hami melon maturity based on the computer vision technology. 6 color components of Hami melon image in RGB and HSV color space were extracted, and the relationship between hue H and Hami melon maturity in HSV color space was studied. The texture features of 6 Hami melon images were extracted by grayscale symbiotic matrix method, and the texture coverage of Hami melon was calculated with two valued texture images. The comparison and analysis of the texture of Hami melon was made. The color features, texture features, color and texture features of the melon images are fused with LDA, PLS-DA, SIMCA and SVM models. The results show that the SVM modeling method is optimal and the single color feature model is superior to the texture feature model, and the SVM model based on the fusion of color features and texture features is the best. The discriminant accuracy of the correction set and the prediction set is 96% and 92%.8. respectively. The Hami melon maturity discrimination model based on the fusion of spectral and image information is established. The results of the Hami melon maturity modeling are compared with the 4 fusion methods of the DS evidence theory, the ELM limit learning machine, the SVM support vector machine and the Ada Boost classifier. The DS evidence based on the decision layer fusion is used. The correction set and the prediction set of the Hami melon maturity discriminant model established by theory are 96% and 92% respectively, which are the same as the optimal discriminant results of a single computer visual color feature model. In the three feature layer fusion model, the correct rate of the correction set is 100%, the SVM support vector machine and the Ada Boost classifier are used to establish the fusion model. The result of the prediction set is 97%, the fusion model of ELM limit learning machine is optimal, and the result of the prediction set is 98%.. The fusion method of the ELM limit learning machine is the best fusion method, the model is the optimal fusion model.9., and the information fusion technology and the single near infrared spectroscopy technology are compared and analyzed, and the computer vision technique is used. The results show that the discriminant result of the Hami melon maturity discriminant model built by hyperspectral imaging technology is better than the near infrared spectrum and the computer vision discriminant model. The result of the computer vision technology is better than the near infrared spectroscopy; the near infrared spectroscopy is better than the near infrared spectroscopy. The model is optimized with the fusion of spectral features and computer vision images. The accuracy rate of the prediction set is 98%, the high spectrum is higher, the near infrared spectrum and the computer vision technology model are improved by 1%, 12% and 8%. respectively. It is proved that it is feasible to distinguish the maturity of Hami melon by spectral and image information fusion technology. The results of this study are at the same time It provides some ideas and methods for Hami melon internal quality prediction and rapid non-destructive testing of maturity. It has certain reference value.
【学位授予单位】：石河子大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：TS255.7

【相似文献】