Research On Nondestructive Detection Of Cerasus Humilis Fruit Quality Based On Spectroscopy And Imaging Technology

Cerasus Humilis fruit is a unique fruit in China,because it is rich in active calcium and easy to be absorbed by human body,also known as ―calcium fruit,‖ the fruit is bright color,unique flavor,rich nutrition,and has the characteristics of high health care and high resistance,so it is known as a ―super fruit‖.As consumers pay more and more attention to the internal quality and external quality of Cerasus Humilis fruit,it is of important significance to explore an effective non-destructive testing methods for Cerasus Humilis fruit quality in order to reduce the testing cost,ensure the quality of Cerasus Humilis fruit,increase the added value of postpartum and product serialization.In this study,Cerasus Humilis fruit was taken as the research object.The internal quality,external defects and maturity of Cerasus Humilis fruit were studied by using visible/near infrared spectroscopy,hyperspectral imaging and information fusion technology,combined with chemometrics and deep learning methods.The main research contents and conclusions were as follows:(1)A quantitative prediction model of spectral information,SSC,firmness and total flavonoids content of Cerasus Humilis fruits were established based on visible/near infrared spectroscopy.The effects of six spectral pretreatment methods(MSC,S-G(5 points),MSC+S-G(5 points),De-trending,MA and MF)on PLSR prediction models of SSC,firmness and total flavonoids content of Cerasus Humilis fruits were compared and analyzed.The characteristic wavelengths of single quality index were extracted by UVE,CARS,RC,SPA,UVE-SPA and UVE-CARS methods,and the linear(PLSR)and nonlinear(LS-SVM,2D-CNN)prediction models were established respectively.The results showed that MSC-UVE-CARS-PLSR model had the best prediction effect on the SSC index of Cerasus Humilis fruit,and its Rp,RMSEP and RPD were 0.8579,0.9059 and 1.8766,respectively.The De-trending-CARS-LS-SVM was the best model for predicting the firmness of Cerasus Humilis fruit,and its Rp,RMSEP and RPD were 0.9092,0.9169 and 2.1485,respectively.The S-G(5 point)-SPA-LS-SVM model had the best performance in predicting the total flavonoids content of Cerasus Humilis fruit,with Rp,RMSEP and RPD of 0.9102,1.0613 and 1.8656,respectively.It provided the research basis for the development of nondestructive testing software for the internal comprehensive quality index of Cerasus Humilis fruit.(2)The method of nondestructive testing of normal and defect samples(rust spot,insect damage and crack)of Cerasus Humilis fruits were studied based on hyperspectral imaging technology.From the perspective of spectral information(945~1675nm),the PLS prediction models established by five pretreatment methods(SG,SNV,MSC,BC and De-T)were compared and analyzed,and the PLS model established by De-T preprocessing of raw spectral data had the best performance.The characteristic wavelengths were extracted by RC,SPA and CARS algorithms,and the linear(PLS-DA)and nonlinear(BPNN,LS-SVM)detection models based on the full spectrum and the characteristic wavelengths were established,respectively,and the normal and defect samples(rust spot,insect damage and crack)were identified.The results showed that the overall discriminant accuracy of LS-SVM based on full spectrum was the highest(88.57%).The LS-SVM based on the characteristic wavelengths extracted by CARS algorithm had the highest discrimination accuracy,and the discriminant accuracy of correction set and prediction set was 86.35% and 91.43%,respectively.(3)An end-to-end convolutional neural network spectral qualitative analysis model(1D-CNN)was proposed and applied to the classification of Cerasus Humilis fruits defects.To verify the effectiveness of the model,the discriminant accuracy of the model was compared with that of the traditional optimal discriminant models(RC-PLS-DA,CARS-BPNN,CARS-LS-SVM),and the optimal 1D-CNN model was obtained,the model for correction set and prediction set was 96.83% and 95.24%,respectively.Through the confusion matrix,the discrimination accuracy of the model for rust spot fruit,crack fruit,insect damage fruit and normal fruit were 91.30%,95.24%,90.48% and 100.00%,respectively,and the average accuracy of classification was 94.26%.(4)From the perspective of image recognition,the detection effects of normal and defect samples of Cerasus Humilis fruits were analyzed.Based on the 13 feature bands selected by CARS algorithm,the PCA and MNF algorithms were used to extract feature images,and comparison showed that the PCA algorithm was more effective in recognition.The Imfill algorithm and the Canny edge detection operator combined with the Regiongrow algorithm and the Bwareopen algorithm were proposed to identify the characteristic region of Cerasus Humilis fruit defect.research showed that the overall recognition accuracy of 91.43% for the prediction set by this image processing algorithm.GLCM and Tamura were used to extract the texture feature parameters of the region of interest of the Cerasus Humilis fruit feature image,and the PCA algorithm was used to reduce the dimension of the normalized texture feature parameters,the LS-SVM model based on the characteristic spectrum,texture features and spectral information fusion texture features was compared to discriminate the defect category of Cerasus Humilis fruit.The comparison showed that the LS-SVM model based on feature spectrum fusion texture features was the best,and the discrimination accuracy of correction set and prediction set were 96.19% and 94.29%.(5)In order to study the prediction effect of hyperspectral imaging technology on the internal quality of Cerasus Humilis fruit at different maturity periods,taking Cerasus Humilis fruit at different maturity periods as the research object.Based on the spectral information of 945~1675nm,the effects of five spectral pretreatment methods(S-G,SNV,MSC,BC,De-T)on linear PLSR and nonlinear LS-SVM prediction models of SSC,firmness and total flavonoids content of Cerasus Humilis fruits were compared and analyzed.The effects of SPA,CARS,UVE,RF,UVE-SPA and UVE-CARS on the prediction results of linear MLR and nonlinear LS-SVM models are compared.The results showed that the prediction performance of the nonlinear LS-SVM model was better than that of linear MLR model under the same algorithm,and the model was more stable.The SPA-LS-SVM model predicted the best results for the Cerasus Humilis fruits SSC,with Rp,RMSEP and RPD of 0.8526,0.9703 and 1.9017,respectively.The UVE-CARS-LS-SVM was the best prediction model for the firmness of Cerasus Humilis fruits,with Rp,RMSEP and RPD of 0.7879,1.1205 and 2.0221,respectively.The SPA-LS-SVM model was the best for the prediction of flavonoids content in Cerasus Humilis fruits,with Rp,RMSEP and RPD of 0.9104,1.9039 and 2.6101,respectively.(6)Spectral information and image features(texture + color)information of the image data had been fused based on hyperspectral imaging technology,and the maturity and total flavonoids content of Cerasus Humilis fruit samples were detected.With characteristic spectrum,image feature,and spectral information fusion image feature as the input variables of qualitative discriminant models(PLS-DA,LS-SVM)and quantitative prediction models(MLR,LS-SVM),the maturity discriminant model and total flavonoids content prediction model were established respectively,and the optimal discriminant and prediction models were optimized by particle swarm optimization(PSO).The results showed that the LS-SVM model established by feature spectrum fusion image features had the best discriminative effect on maturity and total flavonoids content prediction.Among them,the discriminative accuracy of the PSO-LS-SVM model for the maturity of the prediction set was 96.25%,and the Rp,RMSEP,and RPD of the total flavonoids content prediction model were 0.9389,1.8723,and 2.6541.(7)The qualitative discrimination of Cerasus Humilis fruit in different maturity periods by hyperspectral imaging technique was studied.Based on the spectral information of 945~1675nm,the discriminant results of PLS-DA and LS-SVM models established by five pretreatment methods(SG,SNV,MSC,BC,DE-T)for Cerasus Humilis fruits at different maturity periods were compared.The discriminant result of LS-SVM model established by DE-T method was the best(correction set was 87.92%,prediction set was 88.75%),and compared with the raw spectral modeling,the discriminant accuracy increased by 3.34% and 1.25%,respectively.The effects on discrimination models(PLS-DA、LS-SVM)results for Cerasus Humilis fruit maturity base on characteristic wavelengths for total flavonoids content and development time of Cerasus Humilis fruit were studied.Research showed that the discriminant results of model based on the characteristic wavelengths of different maturity periods was better than the model based on characteristic wavelengths of total flavonoids in Cerasus Humilis fruit,and the nonlinear LS-SVM model was more suitable for the discrimination of Cerasus Humilis frui at different maturity periods.The SPA-LS-SVM model showed the best discrimination results,and the number of selected wavelengths(19 wavelengths)only accounted for 8.26% of the full spectrum(230 wavelengths),and the discriminat accuracy of the calibration set and prediction set were 85.00% and 87.50%,respectively.(8)In order to realize the nondestructive detection and visualization of total flavonoids content in Cerasus Humilis fruits during different storage periods,hyperspectral images of Cerasus Humilis fruits at 895~1700nm were collected by hyperspectral imaging technology.The average spectra of the region of interest of each sample were extracted,and the Monte-Carlo outlier detection method was applied to identify four abnormal samples(No.8,61,80 and 207).Based on the wavelength range of 945~1675nm(230 wavelengths)and different pretreatment methods,the PLSR model of total flavonoids content in Cerasus Humilis fruits were established,It was concluded that the prediction result of BC-PLSR model was the best.Five effective wavelength selection algorithms(x-LW,CARS,UVE,UVE-SPA,UVE-CARS)were used to extract the characteristic wavelengths,and the linear MLR and nonlinear LS-SVM prediction models based on full wavelength and characteristic wavelengths were established,respectively.The analysis concluded that the LS-SVM model based on the characteristic wavelengths(9 wavelengths)extracted by UVE-CARS combination algorithm had the best prediction ability and robustness for the total flavonoids content of Cerasus Humilis fruits(Rp=0.9357,RMSEP=2.0107,RPD=2.2809).Finally,the optimal BC-UVE-CARS-LS-SVM model was used to predict the flavonoid content of each pixel in the sample,and the visual distribution map of flavonoid content was established combined with the spatial information of the sample.The results indicated that the HSI technology coupled with chemometric algorithms could be used to predict the changes of the total flavonoids content in Cerasus Humilis fruit during storage periods,which provided a theoretical basis for on-line and real-time monitoring of the quality of Cerasus Humilis fruit during storage and the development of multi-spectral imaging system.