Abstract: In the tobacco manufacturing industry, the accurate classification and composition determination of stem, leaf, expanded and recycled tobacco are the key steps in the formulation of tobacco fiber mixing ratio. The accuracy of this process is directly related to the final quality of the cigarette product, and tobacco is a major challenge in identification detection due to its small size and morphological diversity. At present, there are many problems in the field of tobacco component detection, such as low efficiency and low accuracy. In view of this, this study took YOLOv11 model as the basic detection framework, introduced lightweight ShuffleNetV2 as the network backbone, and integrated BiFPN feature fusion mechanism and CA attention mechanism, and proposed a tobacco type recognition and detection method based on the optimized lightweight YOLOv11 model. The experimental results show that the improved YOLOv11 network model can accurately identify four types of stem, leaf, expanded and recycled tobacco while maintaining the advantages of lightweight model. Specifically, the model has achieved excellent performance of 95.8%、87.5%、87.3% and 81.3% in accuracy rate, recall rate, mAP@.5 and mAP@.5:.95 evaluation indexes, respectively. The process of tobacco type identification and detection proposed in this study not only provides a novel and effective solution for component detection in cigarette production practice, but also explores a potential new path for other product type identification and classification tasks.

Key words: Tobacco components, ShuffleNetV2, BiFPN feature fusion mechanism, CA attention mechanism, Improved YOLOv11

摘要： 在烟草制造业中，精确分类与成分测定梗丝、叶丝、膨胀烟丝及再造烟丝这四大类烟草丝品种，是制定烟草丝混合配比的关键步骤。此过程的准确性直接关联到卷烟产品的最终品质，而烟丝因其微小尺寸与形态多样性，构成了识别检测中的重大挑战。当前，烟丝组分检测领域普遍存在检测效率低下与精确度不足的问题。鉴于此，本研究以最新的YOLOv11模型作为基础检测框架，通过引入轻量级ShuffleNetV2作为网络主干，并融合BiFPN特征融合机制与CA注意力机制，提出了一种基于优化轻量级YOLOv11模型的烟丝类型快速识别与检测方法。实验结果显示，改进的YOLOv11网络模型在保持模型轻量化优势的同时，实现了对梗丝、叶丝、膨胀烟丝及再造烟丝四种类型的精确识别。具体而言，该模型在准确率、召回率、mAP@.5和mAP@.5:.95评估指标上分别达到了95.8%、87.5%、87.3%和81.3%的优异表现，同时本文提出的方法对混合烟丝组分测定的平均相对误差均小于7%。本研究提出的烟丝类型识别与检测流程，不仅为卷烟生产中的组分快速识别检测提供了一种新颖且有效的解决方案，同时也为其他产品类型识别与分类任务探索了一种具有潜力的新路径。

关键词: 烟丝组分, ShuffleNetV2, BiFPN特征融合机制, CA注意力机制, 改进的YOLOv11