Traceability of wolfberry origin: A lightweight step-by-step fusion network combining an electronic nose and a hyperspectral system

A multi-sensor system enables the acquisition of comprehensive and diverse sample information from multiple perspectives, allowing different sensing modalities to complement each other and enhance recognition accuracy. However, the integration of multi-source data also introduces a degree of information redundancy, which necessitates the adoption of effective data processing and fusion strategies. To address this issue, we propose a lightweight step-by-step fusion network(AxisFormer), based on a non-destructive detection platform that integrates an electronic nose and a hyperspectral imaging system, for the classification of wolfberries from six different origins. Gas and spectral data are obtained through the electronic nose and hyperspectral system, respectively, with features extracted and fused using a Lightweight Dual-Path Feature Processing Block (LDPB). Multi-Modal Factorized Bilinear Pooling (MFB) then integrates these features for comprehensive analysis. Compared with other traditional classic models and state-of-the-art models, AxisFormer achieves exceptional classification performance, with 99.07 % accuracy, 98.60 % recall, and 98.71 % precision, using a small number of calculations and parameters. This research not only enables effective origin traceability of wolfberries but also provides an efficient and scientific solution for wolfberry quality detection.