Foreign Body Detection in Semen Ziziphi Spinosae Using Hyperspectral Imaging with Electromagnetic Vibration Feeding

The presence of extraneous matter in Semen Ziziphi Spinosae depresses their pharmacodynamic efficacy, while conventional imaging systems suffer from low throughput and slow frame rates. To address this, we engineered an online platform that couples an electromagnetic vibratory feeder with a line-scan hyperspectral imager to identify both extrinsic foreign bodies (Wood pieces, Stones) and intrinsic foreign bodies (Seed shells, Moldy seeds) introduced during fine processing. First, band-ratio coefficients weighted by signal-to-noise metrics were used to isolate six key wavebands. Pixel-wise reflectance at these bands supplied spectral fingerprints for KNN, DT and CNN classifiers. To further boost accuracy, the corresponding band images were denoised and adaptively segmented, and six morphometric descriptors—maximum length (L), maximum width (W), L/W ratio, circularity, projected area and perimeter—were extracted for every segmented object. Spectral and spatial vectors were concatenated and fed into the same three algorithms. Ten-fold cross-validation showed that the CNN fused model outperformed the traditional schemes, delivering a precision of 95.19%, specificity of 98.75%, recall of 99.00% and overall accuracy of 98.80%. The integrated system therefore provides a rapid, non-destructive solution for quality control of herbal food products. • A novel electromagnetic vibratory feeding system enables real-time hyperspectral imaging of small irregular seeds at 58 ms/frame. • Hierarchical dual-strategy wavelength selection reduces 176 bands to 6 optimized wavelengths with 96.6% data compression. • Spectral–spatial fusion activates deep learning efficacy under extreme dimensionality reduction, achieving 98.80% accuracy. • The CNN fusion model significantly outperforms traditional machine learning with robust statistical reliability. • System throughput reaches 4,800 seeds/hour, demonstrating industrial applicability for pharmaceutical food quality control.