Combining optical spectroscopy and machine learning to improve food classification

Near-infrared spectroscopic data, used for non-destructive product identification, are traditionally processed using multivariate data analysis techniques. However, these methods often cover only a limited variability. We target the development of novel machine learning based algorithm enabling identification foreign objects, in combination with food safety and quality evaluation stream, by combining information from ultraviolet, visible, near-infrared reflection spectroscopy fluorescence spectroscopy. Therefore, we implemented classification scheme cascade individual classifiers both types spectral data. In addition, to ease implementation industrial applications reduce processing time, applied feature selection search, limiting considered illumination detection wavelengths 8. As an illustration our algorithm, present walnuts this paper. The optimal consists first classifier on measurements Extreme Learning Machine second Support Vector Machines. A false negative rate good nuts 5.54% was found, while maximal positive equals 8.34%, shriveled walnuts. All other sample defects, including objects molds, show correct exceeding 98%. Consequently, excellent performance indicates strength multipurpose applications.