Non-destructive detection and classification of antibiotic residues in pork using hyperspectral imaging with an...

Detecting antibiotic residues in meat is critical for food safety and antimicrobial-resistance risk management. A major challenge remains the development of analytical systems that can differentiate structurally diverse antibiotics in meat matrices without destructive sample preparation. This study presents a transformer-based framework integrating hyperspectral imaging with attention-guided inverse weighting to attenuate meat matrix interference for antibiotic classification in pork. Spectral data were acquired in visible-near-infrared (VNIR, 400-1000 nm) and short-wave infrared (SWIR, 900-1700 nm) regions. A two-stage transformer was developed: Stage 1 extracts meat-related attention patterns from individual samples, while Stage 2 applies inverse weighting to suppress these patterns during antibiotic classification. The soft inverse weighting strategy achieved 95% accuracy on SWIR data, a 34% improvement over the baseline transformer. This attention-based approach provides a robust, non-destructive method for multi-class antibiotic classification in meat products.