Deep learning model compression and hardware acceleration for high-performance foreign material detection on poultry 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: 2/4/2025
Publication Date: 2/6/2025
Citation: Khan, Z., Yoon, S.C., Bhandarkare, S.M. 2025. Deep learning model compression and hardware acceleration for high-performance foreign material detection on poultry meat using NIR hyperspectral imaging. Sensors. 25(3):970. https://doi.org/10.3390/s25030970.
DOI: https://doi.org/10.3390/s25030970

Interpretive Summary: A study was conducted to integrate hyperspectral imaging (HSI) technology with deep learning (DL) algorithms for real-time detection of foreign materials in poultry processing. Although HSI-based DL models offer high detection accuracy, the complexity of these models and the substantial amount of data captured by HSI, due to the rapid pace of poultry processing, present significant challenges for practical real-time model inference. To tackle these challenges, the goal was to enhance model inference performance through two strategies: post-training quantization and hardware acceleration. Post-training quantization, specifically using half-precision (FP16), reduced model size and computational requirements without compromising accuracy. Hardware acceleration with Nvidia’s TensorRT further improved inference speed. Simulations using hyperspectral line-scan cameras showed that these optimizations achieved a sevenfold reduction in inference time and halved the inference model size while maintaining high detection accuracy. The results of this study suggest that combining post-training quantization with hardware acceleration techniques can effectively overcome computational bottlenecks. This integration makes real-time inference for foreign material detection using HSI-based DL models feasible in industrial poultry processing environments. Ultimately, this approach contributes to enhancing food safety standards and quality control in the poultry industry.

Technical Abstract: Ensuring the safety and quality of poultry products requires efficient detection and removal of foreign materials during processing. Hyperspectral imaging (HSI) offers a non-invasive mechanism to capture detailed spatial and spectral information, enabling the discrimination of different types of contaminants from poultry meat and fat. When combined with state-of-the-art deep learning models, HSI systems can achieve high accuracy in foreign material detection. However, the high dimensionality of HSI data, the computational complexity of deep learning models and the high-paced poultry processing plants settings present challenges for real-time implementation in industrial settings, where processing speed is critical. In this study, we address these challenges by optimizing deep learning inference for HSI-based foreign material detection through a combination of post-training quantization and hardware acceleration techniques. We applied half-precision (called FP16) post-training quantization to reduce the precision of model parameters, decreasing memory usage and computational requirements without any loss in model accuracy. Additionally, we leveraged hardware acceleration utilizing TensorRT module for Nvidia GPU to further enhance inference speed. We conducted simulations using two potential hyperspectral line-scan cameras to evaluate the feasibility of real-time detection under industrial conditions. The simulation results demonstrated that our optimized models could achieve inference times compatible with the line speeds of poultry processing lines, indicating the potential for real-time deployment. Specifically, we observed a 7 times reduction in inference time and a 50% reduction in model size, while maintaining high detection accuracy. Our findings suggest that the integration of post-training quantization and hardware acceleration is an effective strategy for overcoming the computational bottlenecks associated with deep learning inference on HSI data.