Detection Of Winter Jujube Fruits Based On Machine Vision And Near-infrared Hyperspectral Image

Winter jujube is sweet and delicious with thin skin and flesh crisp.It is rich in vitamin C and minerals and widely favored by the public.Currently,there are two ways of jujube grading including mechanical machine and manpower.Jujubes can be divided into multiple grades efficiently based on size by mechanical machine,but it can't detect disease damage.It is necessary to study the intelligent detection method of winter jujube.There are several aspects of varied needs of jujube grading.According to GB/T 22345-2008 Quality grade of fresh jujube,jujube is divided into four different grades which are unripe,white ripe,crisp ripe,ripe.Not only is the price different for jujube with different maturity,the application value is not the same.Sight damage of jujubes is caused by fall during mechanical vibration picking.it will not only affect the quality of their own fruits,and even makes a large number of jujubes metamorphic.In this paper,the identification of different maturity,slight damage and various disease damage of winter jujube were taken as the research contents,which provided the theoretical basis for the automatic nondestructive detection of winter jujube,the conclusions are as follows:(1)For the detection of jujube maturity,nine color components of RGB,HSB and L*a*b* color models were extracted from images of unripe,white ripe and crisp jujubes.Significantly different variables between groups of jujubes were determined by single factor analysis of variance(one-way ANOVA).Then the most significant color components H and a are obtained through a Fisher's least significant differences(LSD)test.By establishing a Bayes linear discriminant function with cross validation,the recognition rate of the extracted pixels was 97.5%.The B component of HSB color model was used for background segmentation.After denoising,the binary image was obtained by Otsu algorithm calculating the threshold automatically.Removed a small area of impurity and then filled the internal small cavities to achieve the image of extracting fruit region.The Bayes linear discriminant function model was used to traverse the image of fruit extracting region of unripe,white ripe and crisp ripe jujube to get recognition image.The average and mean square variance of proportion of different maturity area were analyzed to get the criteria of maturity.The criteria of different maturity was as follows: jujube with crisp area occupying 30.0% was crisp jujube,jujube with white ripe area occupying 70.0% was white ripe jujube,and jujube with unripe area occupying more than 50.0% is unripe jujube.The correct rate of recognition of unripe jujube is 95.4%,the correct rate of recognition of white ripe jujube is 98.3%,the correct rate of identification of crisp ripe jujube is 97.5%,and the average correct rate of recognition is 97.0%.(2)For the detection of slight damage caused by drop,the ordinary color RGB image is difficult to identify.The near-infrared hyperspectral spectrum of 900~1700 nm was used to image the crisp jujubes dropping from 1.5 m,1 m and 0.5 m.The region of interests(ROI)of damage and normal area were extracted and the average spectral curve of damaged jujubes and normal jujubes was drawn,and it showed that the spectral reflectance of the damaged area was lower than that of the normal surface.As hyperspectral image contains a large number of spectral bands including a lot of redundant bands,three band selection methods which are successive projection method(SPA),correlation based feature selection(CFS)and Consistency were used to extract feature bands,and the same fearue bands are near 1363 nm and 1691 nm.Then three different classification models of k-Nearest Neighbor(k-NN),Naive Bayes(NB),Support Vector Machine(SVM)were used with ten-fold cross validation,the results showed that NB classification obtained the highest classification rate,the correct rates of classification of jujubes dropping from 1.5 m,1 m and 0.5 m by NB were 82.5%,83.8% and 81.6% respectively.Regardless of the difference in drop height,classification accuracy rate was 82.6% by NB.Finally,the correct rate of slight damage identification of jujube was 81.8% by image information fusion.The damage area of jujube dropping from 1.5 m is bigger than 1 m and 0.5 m,indicating that the drop height has a certain impact on the damage degree of winter jujube.For the detection of a variety of disease damage of winter jujubes,detection methods would be carried out in two categories.One is detecting dark spot disease damage such as ring rot,anthrax,daily burns and crack.Due to the existence of obvious black parts of disease site,it could be detected by color difference.The other is the detection of shrinkage disease in obviously different texture characteristics.For the former,the detection method is similar to the above-mentioned maturity detection.The Nine color components of RGB,HSB and Lab color models were extracted from the images of disease damage area and normal surface.The significant difference in the components between groups of jujubes was obtained by variance significance analysis.In order to further reduce the number of color components,a fisher' LSD test was used to get the R component of the RGB,the S component of the HSB,and the b* component of the L*a*b* as the most effective distinguishing color components between the disease area and the normal region.Then a Bayes linear discriminant function was established.The correct rate of pixel classification is 94.2% by cross validation method.According to the HSB color model of jujube image,the B component was easy to obtain the fruit region from background.Then the Bayes linear discriminant function was used to traverse the images of extracting fruit region of disease damage jujubes and normal jujubes after image process.The average and mean square variance of the ratio of disease damage recognition area to fruit area were obtained to get the criteria of discriminating disease damage and normal jujubes.The threshold of ratio of disease damage area to the fruit area was 5%,and the final success rate of classification is 89.6%.For the latter,the gray-level co-occurrence matrixes of normal jujubes and jujubes with shrinkage disease were calculated to obtain five texture characteristic parameters including correlation,energy,entropy,moment of inertia and inverse gap.Then a SVM classification model was established by texture characteristic parameters.The effects of two structural parameters including distance(d)and gray level(L)on texture characteristic parameters were explored.The results showed that when the distance was 1 and the gray level was 32,the detection of shrinkage disease reached the expected effect,and the correct rate of classification was 99.4% with processing time of 10.2s.