Study On Quality Detection Of Crisp Pears Based On Hyperspectral Imaging Technology

As one of the three major fruits in my country,pears are deeply loved by consumers.Hardness is an important indicator of the maturity of pears,and sugar content affects the taste of pears and is also an important parameter to reflect the quality of pears.Slight bruising of pears is generally difficult to detect with the naked eye,which greatly reduces the overall quality of pears.At present,the method of detecting the internal quality of pears is mainly to use traditional physical and chemical determination sampling damage detection,which is time-consuming and cannot meet the needs of large-scale grading;machine vision technology can be used Quality is tested.Therefore,it is of great significance to propose a method for rapid and non-destructive detection of pear internal and external quality.This paper takes crispy pears as the research object,and uses hyperspectral imaging technology to simultaneously detect and analyze the internal and external quality of crispy pears,such as hardness,sugar content,slight bruises and bruise time.The research content is as follows:(1)Collect the spectral information of pears based on hyperspectral imaging technology,and collect the physical and chemical value data of hardness and sugar content of pears.Multivariate Scatter Correction(MSC)and Direct Orthogonal Signal Correction(DOSC)methods were used to preprocess the spectrum,and Competitive Adaptive Reweighting Algorithm(CARS)and Sequential Projection Algorithm(SPA)were used to filter the characteristic wavelengths,and 46 wavelengths were obtained respectively.and 8characteristic wavelengths,in order to explore the influence of modeling methods on the hardness prediction results,three methods of partial least squares regression,support vector regression,and ridge regression were used to establish a quantitative prediction model of pear hardness.The results showed that the DOSC-CARS-Ridge model had the best prediction effect,with Rp2=0.9162 and RMSEP=0.3133.(2)In order to improve the predictive effect of the Support Vector Regression(SVR)model,the Genetic Algorithm(GA)was used to optimize the SVR and model the sugar content of pears.The results showed that the SVR prediction after convolutional smoothing(SG)preprocessing Among the models,the prediction effect of the GASVR model is better than that of the SVR.The RP2(0.8520)of the GASVR model is 0.0685 higher than that of the SVR model RP2(0.7835);the RMSEP(0.4138)is 0.0304 lower than that of the SVR model RMSEP(0.4442).In order to further improve the accuracy of the model,the CARS and SPA methods are used to screen the characteristic wavelengths.The model established with fewer wavelengths has achieved better results than the full wavelength.The GASVR established with the characteristic wavelengths screened by the CARS and SPA methods as input Among the prediction models,the CARS-GASVR model performed best,with Rc2=0.8966,Rp2=0.8711,RMSEC=0.4402,and RMSEP=0.4445.The SVM parameters selected after genetic algorithm optimization are C=3.22,g=0.51.Compared with the fullwavelength GASVR model,the coefficient of determination Rp2 of the prediction set is0.0191 higher than that of the full-wavelength GASVR model,and the root mean square error RMSEP of the prediction set is 0.0307 lower than that of the full-wavelength GASVR model.(3)Using hyperspectral spectrum to obtain the spectral information of crisp pear bruises,using principal component analysis(PCA)to screen the characteristic wavelengths of the preprocessed spectra,and using support vector machine(SVM)to establish the discriminant model for slight bruises of crisp pears and crisp pear bruises respectively.A model for temporal classification of pear bruises.The original spectra were preprocessed,and the discriminant model of light bruises of pears and the time classification model of bruises of pears were respectively established by using the recurrent neural network with improved hidden layer.The results show that the hybrid cyclic neural network model Gru-LSTM established by using the normalized original spectrum is the optimal model for pear bruise discrimination,and its prediction set accuracy rate is 96.36%,which is 2.73% higher than that of the SVM discriminant model.The SVM model using convolutional smoothing(SG)preprocessing combined with PCA dimensionality reduction is the optimal model for the time classification of pear bruises,and its prediction set accuracy rate is 90%.