Detecting the authenticity of two monofloral honeys based on the Canny-GoogLeNet deep learning network combined with three-dimensional fluorescence spectroscopy

To determine the authenticity of honey, a deep learning network based on the Canny-GoogLeNet architecture combined with three-dimensional (3D) fluorescence spectroscopy was established. The canny edge detection algorithm was used to extract 3D spectral data from two distinct monofloral honeys, rape honey and wolfberry honey, as well as adulterated honey samples with corn syrup or other types of honey. The dataset was divided into training (133 samples), validation (33 samples), and test sets (12 samples). The classical GoogLeNet architecture was enhanced by optimizing the inception module in Block 2, applying L2 regularization to the improved fully-connected layer, and implementing a monitoring training network model to reduce overfitting and enhance model robustness. The final model achieved approximately 100 % accuracy on the training set and 93.7 % accuracy on the validation set. These results indicate that integrating 3D fluorescence spectra with a CNN-based deep learning model has significant potential for authenticating honey.