Study On Non-destructive Detection Of ’Kyoho’ Grape’s Quality By Multi-perspective Imaging And Nir Spectroscopy Techniques

Grape fruit, known as one of the most four famous fruit, is welcome to all ages for its bright luster and delicious taste. The planting area of grape ranked only second to the citrus in the world. China is the world’s largest producer of table grape, meanwhile it is also one of important importers, resulting in this phenomenon one of the most important reason is that current low level of the commercialization process of table grapes in our country, especially lack of the effective classification, leading to inadequate supplies of high-grade grape and the uneven fruit. In this paper, kyoho table grape was selected as the research object, and the multi-perspective imaging machine vision and near infrared spectral(NIRS) technologies were employed, combined with image processing and chemometrics methods, to non-invasively inspect the internal and external qualities of kyoho grape respectively, and with these methods to realize the grade of table grape. The main research points were concluded as follows:1) Construction of the multi-perspective projective imaging platform and segmentation of the multi-perspective image of bunch. To expand the shooting visual field of mono-camera and to acquire the maximum information of bunch surface, two pieces of the front-reflected flat mirror were employed to reflect the side surface of the hanging grape bunch. The parameters of position and posture between the mirrors, bunch and camera were optimized, and determined that the imaging distance was 800 mm, the angle between two mirrors was 104°, and the distance between two mirrors‘ intersection point to the centre of bunch is 373 mm. The acquired colored image was pre-processed by channels‘ difference operations to enhance the gray contrast between the foreground and the background, and then a method of local mean-variance threshold segmentation was used to segment the foreground area of clusters. According to the invariant height feature of the hanging bunch, the regions of virtual clusters were affined to the size of real region. As a result, three regions of grape bunch between every 120° interval angle projective were obtained, including one real region and two virtual regions.2) Quantitative and qualitative estimation of bunch?s compactness of table grape were conducted on the basis of multi-perspective simultaneously imaging approach. According to the Organisation Internationale de la Vigne et du Vin 2007(OIV 2007), the compactness of grape bunch was quantized in the artificial evaluation way. The bounding volume of grape bunch was calculated by the accumulation of transverse area, and with the weight of bunch, the bulk density could be obtained to estimate the bulk density of bunch, but low correlation to the bunch‘s compactness only with 0.667. Thus the method of using image features to predict the bunch‘s compactness was proposed. Total 23 features of grape bunch were extracted and normalized as the input, and then three quantitative models, included multiple linear regression(MLR), principal component regression(PCR) and partial least-squares regression(PLS), were comparatively developed to predict the output index compactness. The optimal PLS model was selected with correlation coefficient of prediction(rp) of 0.8481, as well as root mean square error prediction(RMSEP) of 1.2287. The compactness indices can be divided into ’tight’, ’moderate’ and ’loose’ three levels according to the Chinese common description. Linear discriminant analysis(LDA) and back propagation artificial neural network(BPANN), using the components compressed by principal analysis components(PCA), were established, and another LDA and support vector machine(SVM), using the components compressed by PLS, were established. In these four classifiers, the optimal PLS-SVM model was the best to classify the grade of bunch compactness with the classification accuracy of 88.33%.3) Recognition of the overlapped berries and evaluation of the number of berry per bunch was conducted by the method of “bunch contour-polygon approximation iteration- least square circle fitting”. The attributes of grape berries was counted, and it was found that the frequency histogram of berry‘s weight in clutser was usually in the symmetrical distribution, and the shape of grape berry was similar to ellipsoid, whose radius of equivalent circle could be regarded as the berry‘s radius. The contour was extracted by tracking the edges of bunch region in the binary image. Two methods, included ―contour rotation- local extremum‖ and ―polygon approximation iteration‖, were comparatively developed to search out the breakpoints on the bunch contour, and the second method was outperformed for its more number of the right segmented breakpoints and less execution time. The segmental arc between two neighboring breakpoints was fitted by the least squares circle fitting(LSCF), and the eligible circles were screened out by series of limited conditions. Then the weight of berry could be inferred by density-volume equation, and the average weight of berry per bunch was associated to the actual weight, with rp of 0.9033, RMSE of 0.7563 g per berry, as well as the average deviation of 6.98%. Assumed that the screened circles obeyed in the normal sampling distribution in each cluster, the number of berry per bunch was associated to the actual number of berry, with rp of 0.9276, RMSE of 5.563, as well as the average deviation of-4.15%.4) Internal qualities and sensory favor of table grape were inspected on the basis of visible/near infrared spectroscopy(NIRS). The diffuse reflection spectra and transmission spectra were respectively collected the off-cluster berries on the homemade platform, and it was regarded that the transmission spectra could effectively access the internal information of fruit. The best PLS model predicted sugar content(SC), total acidity(TA) with rp of 0.885 and 0.773, as well as RMSEP of 0.598 and 0.048, respectively. Sensory score of berry sample was evaluated by a panel of common consumers, but the scores were lowly correlated to SC and TA with only 0.528. The favor levels of sample‘s sensory preference were divided into two categories according to the sample‘s sensory score. The transmission spectra were compressed by PCA, and principle components were used as the input of classifiers. Three models, including BPANN, SVM and extreme learning machine(ELM) were comparatively developed to classify the sensory favor, and PCA-ELM classifer was the best with corrective rate of 78.7%. A type of negative pressure spectral acquisition in diffuse refection mode was designed to measure the berry‘s interiors on-cluster. Three variable selection methods including uninformative variable elimination(UVE), adaptive weighted sampling method(CARS) and genetic algorithm(GA) were comparatively employed to optimize the PLS model, and the UVE-PLS model was the best to predict the berry‘s SC with rp of 0.6862, as well as RMSEP of 0.9972. By predicting the berries of the cluster, it was found that the value of berry‘s SC per cluster was not formed in the symmetric or normal distribution, with a span range of generally above 5% brix in the same cluster. An evaluation method of cluster‘s SC was suggested to average several berries in different cluster‘s positions.5) Gradation was conducted on the basis of the internal and external qualities of kyoho table grape by multi-perspective imaging and spectroscopy technologies. The grade standard of table grapes was summarized, and in this work five indicators, such as the external qualities shape, color, compactness, average weight of berry, and the internal quality average SC of cluster, were used to grade the table grape. The qualified bunch shape was screened out through the bunch ratio of width to height, weight and parameters of contour‘s curve. The color image of grape region was transfered from RGB to HIS color space, and the riped regions of bunch were segmented out from the Hue channel image. The colored rate of grape bunch was measured by summation of riped area in different perspective bunch region. Overall qualities of grape was classified into four grades by the logic judge method, and 47 samples was repeatedly tested four times, with results of only 16.6% super grade, 24.5% first grade, 23.9% second grade and 39.1% off-grade. The accuracy of the repeated tests for the grade of table grape was 86.9%.This study aimed to enhance the objectivity and timeliness in the evaluation of table grape‘s quality, and provided new methods and ideas for the automatic, rapid detection and gradation of table grape‘s quality. The findings were also of great significance for the scientific guidance of post-harvest process and grade of table grape, increasing the market competitiveness, improving the business efficiency.