Mid-infrared spectra feature extraction and visualization by convolutional neural network for sugar adulteration identification of honey and real-world application

In this study, convolutional neural networks (CNN) strategy combined with mid-infrared (MIR) spectra is applied for sugar adulteration identification in honey. In order to acquire the variations from different classes as well as obtain a better interpretation of the model, a visualization of the CNN algorithm is employed for honey adulteration determination. All kernels are co-acted to explore subtle features by learning different weights and biases. The amplified or attenuated signals transformed by the kernels yield positive or negative effects to the model. Furthermore, the visualized features are implemented for spectroscopy interpretation. The overall results demonstrate that CNN is a promising method with improved accuracy over least squares support vector machines (LS-SVM) and partial least squares discriminant analysis (PLS-DA) especially for the market-obtained samples. Besides, this study supports the theoretical and practical foundation for CNN on the aspects of feature selection, visualization and model interpretation.