Abstract: A methodology for the non-targeted detection of milk powder authenticity using Raman imaging was proposed in the present study. Meanwhile, a novel flow-based discrimination neural network was developed to extract the deep feature of the Raman image of milk powder. Using a combination of possibility distribution transformation and non-volume preserving strategies, feature distribution of authentic milk powder was constructed to distinguish between normal and adulterated milk powder samples. As a result, this method could identify various adulterated samples with an accuracy higher than 97.3%, and the limit of detection was 0.3%. The present methodology was characterized by a wide range of applicability, high precision, convenience and rapidity and could meet the demand of milk powder authenticity detection in practice, which may also provide a new approach for non-targeted detection of the authenticity of other non-homogenous food systems.

Key words: milk powder authenticity; non-targeted detection; flow-based discrimination neural network; Raman imaging