Abstract: This work develops a Raman spectral classification method using convolutional neural network (CNN-Raman) for detecting milk powder adulterants. Using a Raman hyperspectral imaging platform, the raw spectra of sufficient milk powder samples were collected and preprocessed by discrete wavelet transform (DWT). Subsequently, the DWT-filtered spectra were used as the input of CNN to construct a multivariate model. The classification results before and after spectral preprocessing were investigated. Unexpectedly, inappropriate spectral preprocessing worsened the classification performance of the CNN model, while the raw Raman spectra were accurately identified by the CNN. The CNN model based on the raw Raman spectra was capable of identifying an unknown sample accurately with a recognition rate of 95.5%. These results reveal that CNN can be combined with spectral preprocessing and modeling to greatly simplify the calculation process of Raman spectral classification. The CNN-Raman method represents a promising tool for quality and safety inspection of milk powder samples.

Key words: Raman spectroscopy; milk powder adulterants; spectral preprocessing; spectral classification; convolutional neural network