Origin identification of millet by combining spectral detection and a group attention feature calculation and classification network

Abstract

The identification of millet origins serves not only a tool for quality control but also as a bridge connecting producers and consumers, as well as traditional agriculture and modern markets. Spectral analysis captures absorption features associated with various chemical bonds, combination bands of functional groups, and overtones, thereby enabling the characterization of both the chemical composition and physical properties of millet. In this work, we employ spectral analysis technique and propose a deep neural network, termed the Group Attention Feature Calculation and Classification Network (GAFCC-Net), to achieve origin authentication of millet. First, a hyperspectral imaging system is used to acquire spectral data from millet samples, a Savitzky–Golay (S-G) smoothing filter is applied to improve the signal-to-noise ratio, and Standard Normal Variate (SNV) transformation is performed to enhance intra-class data correlations. Next, we propose a lightweight Group-Attention Feature Computation Module (GAFCM), which integrates group convolution with both local and global attention mechanisms for spectral feature computation. Finally, these components are integrated into GAFCC-Net for origin classification. Ablation studies confirm the design validity of GAFCM. Compared with state-of-the-art methods, GAFCC-Net achieves superior performance, yielding an accuracy of 98.86 %, a precision of 98.81 %, and a recall of 98.86 %. In summary, the integration of spectral analysis with GAFCC-Net provides an effective technical solution for the traceability and quality control of millet.