Parameter- and Compute-Efficient Spatial-Spectral Transformer Framework for Pixel-Level Classification of Foreign...

Area of Science:

• Food Science and Technology
• Artificial Intelligence
• Spectroscopy

Background:

• Foreign plastic objects (FPOs) in poultry pose significant food safety risks and economic challenges.
• Conventional detection methods like X-rays and color imaging are insufficient for small or low-density plastics.
• Hyperspectral imaging (HSI) offers rich data but faces computational challenges in industrial settings.

Purpose of the Study:

• To develop a computationally efficient spatial-spectral transformer framework for pixel-level classification of FPOs in broiler meat using near-infrared HSI (1000-1700 nm).
• To improve the speed and accuracy of FPO detection in industrial poultry processing environments.

Main Methods:

• Introduced a parameter-efficient spatial-spectral transformer framework integrating center-focused linear attention (CFLA), patch-local mixed-axis 2D rotary position embedding, and low-rank factorized projection (LRP) matrices.
• Trained and evaluated the model on a dataset of 52 chicken fillets with labeled hyperspectral pixels from 12 polymer types.
• Utilized NIR-HSI data for pixel-level classification of FPOs.

Main Results:

• Achieved 99.39% overall accuracy, 99.57% average accuracy, and a 99.31 Kappa coefficient on test data.
• Demonstrated significant efficiency gains: 83% reduction in MACs, 22% fewer parameters, and a 4.19x speedup over the baseline.
• Exceeded 98.05% precision, 98.59% recall, and 98.76% F1-score across all classes.

Conclusions:

• The proposed framework effectively combines high classification performance with computational efficiency for real-time FPO detection in poultry.
• Enables high-throughput inference, supporting contamination source traceability and preventive quality control in industrial poultry processing.
• Sets a benchmark for applying transformer-based models to food safety inspection tasks.