Qualities Analysis Of Fresh Ziziphus Jujube Cv.junzao In Akesu Based On Near-infrared Spectroscopy And Machine Vision Technology

Nutrient composition of Ziziphus jujube cv.junzao is rich.Jujube is also an important economic crop in Xinjiang.Sorting fresh Ziziphus jujube cv.junzao is conducive to remove the defective products before processing it and can save energy consumption and cost of drying.So,this article is based on the near-infrared spectroscopy,machine vision technology,and information fusion technology,it is conducted for the prediction of size,quality and sugar content of fresh Ziziphus jujube cv.junzao produced in Akesu of Xinjiang,as well as the algorithm models,were wrote for sorting cracked fruit,size and grade of fresh Ziziphus jujube cv.junzao.The key findings were as follows:(1)Weight grade sorting.The size,texture and chrome properties of fresh Ziziphus jujube cv.junzao are obtained by cutting original image,and processing cutting image with the median filter,and binarization of preprocessed image with the threshold of Otsu algorithm,and post processing of the binary image.The correlation coefficient is 0.9920,and the relative error is 0.0259.2 when the first order linear regression was used for predicting weight of Ziziphus jujube cv.junzao with four features of area,long axis,short axis,and circumference;The total accuracy was 95.45% for size grading of fresh Ziziphus jujube cv.junzao by using discriminate analysis and the longitudinal feature of the target region in the binary image of Ziziphus jujube cv.junzao.(2)Crack Recognition.Using the five texture features of entropy,correlation,contrast,energy and smoothness based on gray covariance matrix,the grading accuracy rate of cracked Ziziphus jujube cv.junzao reach up to 89.92% when new five texture features,which were obtained from the high-frequency or low-frequency images after the decomposition of wavelet transform,were used to build the grading model of support vector machine,Which is higher than the recognition rate of 77.28%of the original image gray level co-occurrence matrix.(3)Prediction of sugar content of fresh Ziziphus jujube cv.junzao based on near infrared spectroscopy.First,exception values of the near-infrared spectral and the total sugar of fresh Ziziphus jujube cv.junzao were eliminated through using the Mahalanobis distance and the residual concentration method.Second,the moving window smoothing was determined as a better preprocessing method by comparing the sugar prediction of some pretreatment methods of the spectrum,such as moving window smoothing,data centering,and derivative.Third,evaluating the effect of the four wavelength selecting methods,which include correlation coefficient,interval partial least squares,backward interval partial least squares and genetic algorithm,the results shown that the predictive correlation coefficient is 0.8336,and the root mean square error is 1.89 when the prediction model with wavelength selection of genetic algorithm + partial least squares for predicting the total sugar of fresh Ziziphus jujube cv.junzao is better than that of other models.(4)Grading of fresh Ziziphus jujube cv.junzao based on images and spectral fusion information.The near infrared spectrum and the image information are merged into the feature layer,different extraction pretreatment of spectral feature and image feature using principal component analysis(PCA)extract the feature,and the BP neural network and the support vector machine pattern recognition method are used to establish the sorting model of fresh Ziziphus jujube cv.junzao respectively.The results shown that the 3principal component about machine vision and the 9 principal component feature about near infrared spectroscopy after smooth derivation,the establishment of the comprehensive quality of fresh jujube sorting support vector machine prediction based on recognition rate is 96.92%;the comprehensive quality of fresh jujube sorting established more than BP neural network fusion prediction set recognition higher than the recognition rate of 94.46%;near infrared spectroscopy was 87.69% higher than that of the single sensor;machine vision recognition rate of single sensor 83.09%.