| Hami melon is a general designation of the thick skin melon,whicn is a specialty in Xinjiang.Hami melon is favored by consumer for its sweet and crispy taste,high nutritional and medicinal value.It has been known as “the Chinese first honeydew melon”.With improvement of the people living standards,the internal and exterior qualities of Hami melon have received the unprecedented attention.The traditional methods for Hami melon internal quality are destructive,inefficient and time-consuming,and the method of determining Hami melon maturity mainly rely on subjective experience judgments,which are hard to be highly accurate and fail to satisfy the need of modern production.Focusing on these issues,we chose the different maturity of Hami melon as the research object,the near infrared spectroscopy,hyperspectral imaging technology and computer vision technology were used to detect internal qualities and maturity of Hami melon simultaneously.The feature extraction method of Hami melon spectra and images,and feature fusion method for Hami melon maturity were studied,and Hami melon maturity discrimination fusion model was established at the same time.The main research contents were as follows:1.Analysis on the correlation between Hami melon internal qualities(soluble solids,total acid,firmness)and maturity.The Hami melon maturity is a comprehensive evaluation results with change of various physicochemical indexes during the fruit development.In order to find the most indicator for reflecting maturity of Hami melon,the correlation and one-way ANOVA analysis between soluble solids content,total acid content,firmness of Hami melon and maturity was made in the study.SSC had the best correlation with the maturity of Hami melon,and the SSC had the greatest difference among different maturity melons,which provided a theoretical basis for the selection of appropriate characterization factor of Hami melon maturity.2.The prediction model of the internal quality of Hami melon was established based on near infrared spectroscopy.The effects on PLS and nonlinear SVR prediction models of different spectral pretreatment method such as normalized(Norm),standard normal variate(SNV),multiple scatter correction(MSC)and Savitzky-Golay smoothing(SG-1-Der)were studied.Research showed that the better pretreatment methods were Norm and MSC.The important characteristic wavelength must be selected in order to simplify the model and improve the prediction ability of the model,uninformative variable elimination(UVE),genetic algorithm(GA)and successive projections algorithm(SPA),and competitive adaptive reweighted sampling algorithm(CARS)were used to select characteristic wavelength of SSC,TAC and firmness of Hami melon,and establish the PLS and nonlinear SVR prediction models of SSC,TAC and firmness of Hami melon by using different characteristic wavelength selection methods.Research showed that The PLS model is better than SVR model for determining the SSC,TAC and firmness of Hami melon.The GA method was the best characteristic wavelength selection methods,and was used to select effective characteristic wavelengths for the SSC,TAC and firmness of Hami melon,respectively.The GA-PLS model achieved optimal prediction for SSC with RP,RMSEP and RDP values of 0.9061、0.4630 and 2.363,respectively,and 0.8277、0.0391 and 1.781 for TAC,respectively,and 0.8729、28.47 and 2.049 for firmness,respectively.3.The discrimination model of the maturity of Hami melon was established based on near infrared spectroscopy.The effects on discrimination results for Hami melons maturity by linear discriminant analysis(LDA),partial least squares discriminant analysis(PLS-DA),soft independent modeling of class analogy(SIMCA)and support vector machine(SVM)with different spectral pretreatment method(Norm,SNV,MSC,SG-1-Der)were studied.Research showed that the better pretreatment methods were SNV and MSC.The SVM model was the best than linear LDA,PLS-DA,SIMCA model for discriminating the maturity of Hami melon.The discriminant results of calibration and prediction set for SNV-SVM model were 94% and 88%,respectively.The effects on discrimination results for Hami melons maturity base on characteristic wavelength for SSC and maturity of Hami melon were studied.Research showed that the discriminant results of model base on maturity characteristic wavelength was better than the model base on SSC characteristic wavelength of Hami melon.The obtained optimal numbers of wavelength variables for maturity were 7 by SPA,accounted for only 2.73% of full spectral variables,which was 914.75 nm,921.26 nm,934.28 nm,940.79 nm,1318 nm,2132.9 nm,and 2364.8 nm,respectively.The selected variables by SPA were used as the inputs of SVM model to build the best model for discriminanting the maturity of Hami melon.The discriminant results of calibration and prediction set for SPA-SVM model were 93% and 86%.Compared with the full spectral SNV-SVM model for maturity,the discriminant results of calibration and prediction set were only decreased by 1% and 2%.4.The prediction model of the internal quality of Hami melon was establish based on hyperspectral imaging technology.The effects on PLS and nonlinear SVR prediction models of spectral pretreatment method such as Norm,SNV,MSC and SG-1-Der were studied.Research showed that the better pretreatment methods were SNV and MSC;the SNV-SVR model performed better than the MSC-PLS model for SSC prediction;For TAC and firmness prediction,compared with the PLS model,the predictive results of SVR were increased by 1% and 2%.The important characteristic wavelength must be selected in order to simplify the model and improve the prediction ability of the model.Synergy interval partial least squares(si-PLS),GA,SPA and CARS were used to select characteristic wavelength of SSC,TAC and firmness of Hami melon,and establish the PLS and nonlinear SVR prediction models of SSC,TAC and firmness of Hami melon by using different characteristic wavelength selection methods.Research showed that the CARS method was the best characteristic wavelength selection methods,and was used to select effective characteristic wavelengths for the SSC,TAC and firmness of Hami melon,respectively.The SVR model was better than PLS model for predicting the SSC,TAC and firmness of Hami melon.The CARS-PLS model achieved optimal prediction for SSC with RP,RMSEP and RDP values of 0.9394,0.4071,2.917,respectively,and 0.8705,0.0322,2.032 for TAC,respectively.For firmness prediction,compared with the full spectral MSC-SVR model,the predictive results of model base on characteristic wavelengths was the worst.5.The discrimination model of the maturity of Hami melon was established based on hyperspectral imaging technology.The effects on discrimination results for Hami melons maturity by LDA,PLS-DA,SIMCA and SVM with different spectral pretreatment method(Norm,SNV,MSC,SG-1-Der)were studied.Research showed that the better pretreatment methods were SNV,and the SVM model was the best than linear LDA,PLS-DA,SIMCA model for discriminating the maturity of Hami melon.The discriminant results of calibration and prediction set for SNV-SVM model were 98% and 94%,respectively.The effects on discrimination results for Hami melons maturity base on characteristic wavelength for SSC and development time of Hami melon were studied.Research showed that the discriminant results of model base on maturity characteristic wavelength was better than the model base on SSC characteristic wavelength of Hami melon.The obtained optimal numbers of wavelength variables for maturity were 57 by CARS.The selected variables by CARS were used as the inputs of SVM model to build the best model for discriminanting the maturity of Hami melon.The discriminant results of calibration and prediction set for CARS-SVM model were 100% and 95%.The CARS method could select the key characteristic wavelengths that were related to Hami melon maturity,which be to not only simplifie the model but also improve the discriminant result.6.In the study of the characteristic wavelength selection from hyperspectral image data,8 characteristic wavelengths were selected by SPA from 57 characteristic wavelengths by CARS,which was 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.Texture features of characteristic wavelength images were extracted from each characteristic wavelength images by grey-level co-occurence matrix,the discriminant model of Hami melon maturity were built based on texture features,characteristic wavelength and data-fusion,respectively.Research showed that the result of spectral model was better than image model,and the result of SVM fusion model was best base on spectrum and texture feature.The discriminant results of calibration and prediction set for fusion SVM model were 100% and 97%.7.The discrimination model of the maturity of Hami melon was established based on computer vision technology.The six color features were extracted from Hami melon image in RGB and HSV color spase,and the relationship between the hue in HSV color space and Hami melon maturity was studied.The six texture features of Hami melon image were extracted by grey-level co-occurence matrix,and texture coverage of Hami melon was calculated from the binary image.The effects of different methods(LDA,PLS-DA,SIMCA and SVM)on modeling results for Hami melon maturity are compared.Research showed that the result of SVM model was the best than LDA,PLS-DA,SIMCA model.The result of model base on color feature was better than the model base on texture feature.The result of SVM model was the best base on fusion of color and texture feature,the discriminant results of calibration and prediction set for fusion SVM model were 96% and 92%,compared with the color feature model,the discriminant results was increased by 1%.8.The discrimination model of the maturity of Hami melon was established based on fusion of near-infrared spectrum and image.The effects of different fusion methods(DS evidence theory,Extreme learning machine,Adaboost classifier and SVM)on modeling results for Hami melon maturity were compared.In decision level fusion,the discriminant results of calibration and prediction set for fusion DS model were 96% and 92%,which was same with that of color feature model.In feature level fusion,the discriminant results of calibration set for three fusion model were 100%;the discriminant result of SVM fusion model was 97%,which was the same with that of Adaboost fusion model.ELM fusion model is optimal,and the discriminant result is 98%.The result showed the ELM fusion model is optimal for Hami melon maturity in information fusion models.9.The discrimination results for Hami melons maturity by single near infrared spectroscopy,hyperspectral imaging technology or computer vision technology was compared with information fusion technology.The result showed the discrimination results for Hami melons maturity based on hyperspectral imaging technique was better than that base on near infrared spectroscopy or computer vision technology;The discrimination results for Hami melons maturity based on computer vision technology was better than that near infrared spectroscopy.The discrimination results for Hami melons maturity based on fusion of near infrared spectrum and computer vision image was the optimal,and the discriminant result was 98%.Compared with the hyperspectral imaging,near infrared spectrum and computer vision model for Hami melons maturity,the discriminant results of fusion model were increased by 1%,12% and 8%.It is feasible to discriminate Hami melon maturity by spectral and image information fusion technology.The research offers a new idea to assess internal qualities and maturity of Hami melon base on information fusion technology simultaneously and there is the referable for non-destructive detection of Hami melon quality. |
PDF Full Text Request