Application of an electronic tongue and hyperspectral imaging with a CNN-transformer fusion model for rapid detection of botanical origins of honey.

Abstract

The botanical origin of honey significantly impacts its nutritional composition, quality, and price. Traditional identification methods are often complex, require expensive equipment, and are time-consuming. This article proposes a rapid detection method for the botanical origin of honey based on an electronic tongue (ET) and hyperspectral imaging (HSI) combined with a CNN-transformer fusion model. First, gustatory and spectral data of honey samples from different botanical origins are collected by ET and HSI systems, respectively. A CNN-transformer fusion model is proposed to perform feature extraction, information interaction, and pattern recognition on the collected ET and HSI data. This model employs a dual-path CNN-transformer to capture local and global features of ET and HSI signals at different scales. A multi-scale interaction module is designed to enhance cross-modal communication and facilitate information sharing between the ET and HSI information. Finally, the contrastive information bottleneck (CIB) module is adopted to optimize mutual information through contrastive learning and enable the integration of ET and HSI features for classification and recognition. The experimental results demonstrate that this method achieves superior recognition accuracy in classifying and identifying honey botanical origin compared to that using either the ET or HSI alone. Its experimental mean test set accuracy, precision, recall, and F1 score reached 99.08%, 99.09%, 99.05%, and 0.9906, respectively. This study provides a new detection method for the botanical source of different kinds of honey, which has a promising application in honey and other food industries.