Hyperspectral Imaging for Honey Quality Detection using Siamese Neural Networks

Honey is a nutritious natural food product with many health benefits and is thus widely utilized as a natural sweetener or consumed as a dietary ingredient. Different botanic origin honey types have various quality, flavor, or health benefits. Therefore their market values differ significantly. Many studies have been devoted to investigating honey quality with various chemically-based techniques. Nevertheless, these methods are expensive, laborious, and time-consuming. In addition, it is impossible to collect honey samples containing all the wide variety of botanical origins or adulteration methods. Thus, a more feasible approach is to develop a databank including authentic honey types of interest, whose data is also easy to process and collect, then designe a model to tell whether an unknown sample belongs to the same class of samples in the databank or not. This paper proposes a new approach using Siamese neural networks designated to learn similarities between hyperspectral imaging of honey samples. Siamese neural networks learning allows models to make correct predictions, given only a single example of a new class. With convolutional neural network architecture, the learned features acquired generalized the discriminating power to predict new unseen images correctly. The average validation accuracy rate we achieved is 95%. We interestingly found that the spectra properties of honey types collected from the same botanic origin produced by different producers vary significantly