Research On Classification Method Of Winter Jujube Based On Vis/NIR Hyperspectral And Machine Vision Technology

Winter Jujuba is a kind of fresh jujuba with good quality.It has thin skin and flesh,delicious flesh,and it is rich in nutrition.It can prevent cancer and prevent cardiovascular and cerebrovascular diseases.Its nutritional value is the highest among with many fruits.In recent years,the output of winter jujube in China has continuously increased,and it has continued to grow in the market.Winter jujuba has been playing a more and more important role in the sustainable development of agriculture and increasing farmers'income.Maturity(coloration degree),soluble solids(Sugar content),the weight of fruit(size)are important indicators of the quality of winter jujube according to GB/T22345-2008?Fresh Jujube Quality Grade?,GB/T32714-2016?Winter Jujube?National Standard and NY/T2860-2015?Winter Jujube Grade Specification?Ministry of Agriculture standards and related works.Due to the differences in illumination,nutrients and development of jujube,etc.,different maturity and size of jujube fruit coexist at the time of harvest,which makes the maturation and size of winter jujube(harvested mechanically or artificially)mixed together,and there are differences in individual taste.It is hardly to realize the sale of jujube fruit according to the quality of value without classification,and labor grading by hands is labor-intensive and time-consuming.Therefore,it is necessary to develop a method that can realize automatic grading of jujube maturity,soluble solids,and single fruit weight.1.Winter jujube maturity test:The hyperspectral information of three grades of mature winter jujube of special/first-class,second-class,and third-class grade were obtained,respectively,and the SG smoothing algorithm was used to denoise the spectral data,and then successive projections algorithm(SPA)and Competitive Adaptive Reweighted Sampling(CARS)and Random Frog(RF)were used to select characteristic wavelengths(CWs)to reduce the spectral dimension.Partial least squares discriminant analysis(PLS-DA)model were established based on the characteristic wavelengths selected by the three algorithms respectively,and the grading effects of three models were compared.The results show that the accuracy of PLS-DA model based on the characteristic wavelengths selected by SPA,CARS and RF is 90.90%,92.72%and 93.63%,respectively.2.Prediction and classification of soluble solids content(SSC).The hyperspectral information of mature winter jujube were collected and the corresponding SSC content of winter jujuba was measured.The hyperspectral information in the range of 415 nm~1000 nm of jujube was used and SG smooth and MSC were used to reduce its noise.CARS-SPA and SPA-SPA were used to selected characteristic wavelengths of soluble solids content;The PLSR models based on full spectra,the selected wavelengths of by CARS-SPA series method and SPA-SPA series method were compared.In addition,the model's practical effect was tested by combining the model with the requirements of the SSC national standards for winter jujube,and the graded accuracy was used as a criterion to evaluate the practical performance of the model.PLSR of full spectra establishment:R_c=0.9613,RMSEC=0.5032,R_p=0.9396,RMSEP=0.6041,rating accuracy:91.57%,the PLSR established by CARS-SPA CWs:R_c=0.9062,RMSEC=0.601,R_p=0.9257,RMSEP=0.675,the classification accuracy rate is89.47%,and the PLSR established by SPA-SPA CWs:R_c=0.9353,RMSEC=0.6985,R_p=0.9165,RMSEP=0.7132,the classification accuracy rate is 86.31%.3.Winter Jujube single fruit weight detection and classification.Winter jujube RGB images and fruit weight information were collected and K-means clustering segmentation algorithm was used to segment jujube images,and they were smoothed by median filter after they were binarized.Then winter jujube binary image size information were extracted and analysised.The contour area is used to establish a fruit weight prediction model.By calculating the number of pixels in the enveloping area of the binary image contour of winter jujube,the actual area of the winter jujube is obtained by converting the proportion of pixels with a knowned area marker.Single fruit weight prediction models for primary red fruit and half red-red fruit were established,respectively.The practical performance of fruit weight prediction models of two kinds of winter jujube were verified reference to the national standard of?Winter jujube?.The correlation coefficient between the contour area and the fruit weight in the prediction model of primary red fruit reached 0.9323 and the accuracy of the model weight classification of the test set was 86.66%.To half red-red fruit the coefficient reached 0.9203,and the precision of model's fruit weight classification was 84.44%,which verified the reasonableness and feasibility of fruit weight classification with jujube fruit size.The results show the combination of Vis/NIR hyperspectral and Machine vision can predict and classify the maturity,SSC and fruit weight,and that can provide a technical reference for the development of automatic classification equipment for winter jujube.