A lightweight neural network for online real-time non-tobacco-related material instance segmentation in multispectral video

The motivation for this research stems from the critical need within the tobacco industry to enhance the quality control and efficiency of flue-cured tobacco production. The inadvertent inclusion of non-tobacco related materials (NTRM) during the production process has been a persistent challenge, significantly affecting the final product's quality. Existing detection methods have proven insufficient in addressing this issue comprehensively. To address this gap, we embarked on the development of a dedicated NTRM detection method tailored to the specific requirements of the tobacco industry. We first established an extensive NTRM spectral database through the utilization of multispectral and visible light imaging techniques. Subsequently, we propose an enhanced YOLO model, called YoRMD-Net, which enables swift and precise NTRM segmentation. Our method achieves a 97.5% mAP in identifying NTRM and offers real-time tracking at 16.5 ms per frame, significantly outperforming traditional manual detection in actual production. By doing so, we aim to provide an effective solution that not only improves product quality but also streamlines production processes, ultimately benefiting the entire tobacco industry by ensuring the production of premium-grade flue-cured tobacco products.