Non-destructive Detection For Insect-infested And SSC Of Changzao Based On NIR Hyperspectral Imaging Technique

Fruit quality is an important factor which determine commodity value of the fruits, and an important attribute that influences the buying requirement of consumers.The traditional physico-chemical detection methods cannot satisfy the needs of fruit commercialization, because of a single-index for detecting, time-consuming and samples which is destructive. NIR Hyperspectral imaging(HSI) technique integrates the advantages of machine vision with near infared spectrum,which can obtained synchronously spectrum information in continuous space, it can visualization research for the internal and external quality of agricultural products and use widely in the detection of fruit quality in recent years.Based on ’LingwuChangzao’ jujube of Ningxia as the research object, this paper developed identification algorithms of insect-infested jujubes using near infrared hyperspectral imaging technology, combined chemometrics methods with image processing method, its quantitative analysis model of soluble solids content of jujubes was established and realized comprehensive evalution of the internal and external quality of jujubes, the detection algorithms offered theoretical basis for developing a real-time, fast and on-line nondestructive detection system. The main research results are as follows:(1) The identification algorithm of insect-infested jujubes based on PCA of optimal wavelengths, combining NIR hyperspectral imaging system with five image processing algorithms, including mask, image negation, threshold segmentation, dilation and connectivity analysis was developed, and240samples were identificated one by one, identification rate of insect-infested jujubes was81.9%, intact jujube was correctly identified96%.(2) Unidentified jujubes were further classified by band ratio (BR) and image subtraction (IS) algorithm with image processing algorithms. The experimental results shows that PCA and IS algorithms achieved identification rate from81.9%to92.5%; PCA and BR achieved identification rate from81.9%to90.6%.(3) Leverage value, compled Cook distance with studentized residual was adopted to eliminate abnormal samples for obtaining hyperspectral data and the actual measured value of soluble solid content (SSC). The robustness of prediction model was improved effectivly.(4) This study adopted multiple scattering correction (MSC) to spectral preprocessing, the optimal wavelength was obtained by the correlation coefficient method, The prediction model of soluble solids content were built by back-propagation neural networks (BP-ANN) and multiple linear regressions (MLR). The results shows that the prediction models of soluble solids content using back-propagation neural networks (BP-ANN) was better than MLR, the correlation coefficient and root mean square error of prediction were0.957and0.9264°Brix, respectively.