Research The Brix Detection Method Of Hetao Melon Based On NIRS And Machine Vision Information Fusion

On the basis of analyzing the domestic and international development status of fruit non-destructive testing technology such as visible-near infrared spectroscopy and machine vision, this paper introduces the online detection system for honeydew melon internal quality, which combines diffuse transmittance spectra and image information. Based on the detection system, the research group explores the rapid non-destructive testing methods for the sugar content of Hetao honeydew melon. The main research work and achievements are as follows.(1) The online detection system for honeydew melon quality which combines visible-near infrared spectroscopy and image information is programmed. After installation and debugging, this system can come to a better realization of the collection of visible-near infrared spectroscopy and image online.(2) Based on the detection system, the Static and online data of the diffuse transmittance spectra and image information for three kinds of Hetao honeydew melons (0.07m/s、0.09m/s、0.11m/s) are collected. On this basis,"Golden Ruby" is taken as a research object, whose sugar content is detected.(3) Visible-near infrared spectroscopy is pre-processed for154samples, by which the characteristic wavelengths of the honeydew melons are extracted and the characteristic value of the transmittance principal components is analyzed by’spss.(4) For the image features of the honeydew melon, the image processing algorithm is used to extract the color, size, and fruit shape index information. The color features are analyzed to find out the principal components which characterize the color features.(5) Two kinds of modeling methods are studied, namely BP neural network algorithm and SVM algorithm. The principal components value of the color characteristics as input and sugar content as output, a model is built. BP neural network training set fitting degree is70%, predicting outcomes of r and RMSE are0.4782and1.8054. SVM algorithm predicts outcomes of r and RMSE are0.5059and1.5949, the regression of which is better than BP neural network.(6) By SVM (Support vector machine) modeling method, the image, spectral information and sugar content are modeled respectively. The effect of the color feature model is the worst, r and RMSE of which are only0.5059and1.5949, having a low correlation. Compared with spectral component model, spectral signature model is better, r and RMSE of which can reach0.7501and1.2058. Regression model combined by color feature and spectral feature, whose r and RMSE can reach0.8473and0.9684, is better than the one built by single color feature or spectral feature. Color feature, size, and shape index as an external feature, and combined with spectral signature, the effect of the model will be the best, r and RMSE of which can reach0.8630and0.9407.The results show that the fusion of the spectral and image information can not only reduce the impact on the spectrum detection accuracy by fruit size and shape differences, but also improve the detection accuracy with the relevance of color information and quality of the honeydew melon.