Parameter and compute-efficient spatial-spectral vision transformer framework for pixel-level classification of foreign plastic objects on broiler meat using NIR-hyperspectral imaging

Authors: KHAN, ZIRAK - University Of Georgia; Yoon, Seung; BHANDARKARE, SUCHENDRA - University Of Georgia

Submitted to: Sensors
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 4/13/2026
Publication Date: 4/16/2026
Citation: Khan, Z., Yoon, S.C., Bhandarkare, S.M. 2026. Parameter and compute-efficient spatial-spectral vision transformer framework for pixel-level classification of foreign plastic objects on broiler meat using NIR-hyperspectral imaging. Sensors. 26(8), 2459.
DOI: https://doi.org/10.3390/s26082459

Interpretive Summary: Foreign plastic materials that accidentally end up in chicken meat products pose significant risks to food safety and can negatively impact the poultry industry economically. Therefore, it would be beneficial for the industry to implement objective sensing and sorting methods that enable the quick and accurate detection and identification of different types of plastics in real-time during production. This study introduces a novel technique using a specialized artificial intelligence framework designed to enhance the efficiency and effectiveness of identifying foreign plastics in chicken meat through near-infrared hyperspectral imaging. Accurate identification of the type of plastic will assist poultry processors in tracing the sources of contamination, such as conveyor belts and employees’ gloves. The proposed framework features computationally efficient and accurate prediction modules. It was tested using a comprehensive dataset that included 52 hyperspectral images of chicken breast fillets and 12 common types of plastics used in packaging and processing in the poultry industry, including materials like PET, PP, and PVC. The new model achieved an overall classification accuracy of 99.43%, correctly identifying most types of plastics present in the chicken samples. Moreover, the efficiency of the new framework showed substantial improvements in computation, reducing the number of calculations required by 83% and decreasing the overall size of the model. This makes it much faster and easier to use in real-world scenarios. In summary, this innovative technique enhances the speed and accuracy of identifying foreign plastics in poultry products, ensuring food safety. Additionally, it offers a practical solution for real-time contamination monitoring, helping industries trace contamination sources and maintain quality control.

Technical Abstract: Foreign plastic objects (FPOs) in poultry products present significant food safety risks and economic threats to the industry. While hyperspectral imaging (HSI) has become a promising non-destructive detection technology, current HSI methods face significant challenges in quickly and accurately detecting and identifying various types of FPOs for real-time industrial applications. These challenges arise from the computational inefficiency of large deep learning models and the insufficient integration of spatial-spectral features inherent in hyperspectral images. This study introduces a computationally efficient spatial-spectral vision transformer framework designed for pixel-level type classification of FPOs on broiler meat, utilizing near-infrared hyperspectral imaging (NIR-HSI). The framework incorporates three key innovations: (1) Center-Focused Linear Attention (CFLA) mechanism that reduces computational complexity from O(n²) to O(n) by calculating attention only for the center pixel while preserving the spatial-spectral context from surrounding pixels, without compromising classification accuracy, (2) Patch-local mixed-axis 2D rotary position embedding (2D RoPE) that preserves the inherent 2D spatial geometry and captures diagonal relationships within hyperspectral patches, and (3) Low-rank projections that achieve approximately 50% reduction in the parameters of transformer projection weight matrices. Our framework was evaluated using a comprehensive in-lab dataset that included 52 hyperspectral images of broiler breast fillets and 12 different types of polymers commonly found in poultry processing environments. The polymers used in the study included packaging materials, such as PET, PP, PS, PVC, LDPE, and HDPE, as well as components of processing and personal protection equipment, which encompassed materials like PUR, ABS, RUB, FAB, NYL, and TEF. The proposed framework demonstrated outstanding classification performance, achieving 99.43% overall accuracy (OA), 99.60% average accuracy (AA), and 99.37% Cohen’s kappa coefficient (Kappa) across 295,340 labeled target pixels within the independent test set derived from these 52 hyperspectral images. Per-class performance also showed consistently high precision (98.11-99.80%), recall (98.70-100.0%), and F1-scores (98.81-99.89%) for all 12 FPO types as well as for the additional chicken meat class, demonstrating robust discrimination even for spectrally similar FPOs. The computation efficiency metrics revealed substantial improvements, with 83% reduction in multiply-accumulate operations (MACs) from 22.51M to 3.80M, 83% reduction in floating-point operations (FLOPs), and 22% reduction in trainable parameters from 451.2K to 352.9K and the model size from 1.72 MB to 1.35 MB compared to the vanilla transformer approach. The proposed framework not only enhances computational efficiency but also demonstrates superior classification accuracy, setting a new benchmark for food safety applications in FPO classification. It provides a practical solution for real-time contamination monitoring that enables source traceability and preventive quality control measures.