Combining optical spectroscopy and machine learning to improve food classification

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