Advanced Tobacco Foreign Object Detection via Optimized Network Architecture and Enhanced Multi-Scale Feature Integration

Abstract The presence of impurities in finished tobacco leaves significantly affects the stability of cigarette quality during the packaging process.To solve the key technical problems that after the cutting process, these impurities become small in size and are similar in shape to the tobacco leaves, making accurate detection difficult. This paper proposes a method for detecting foreign objects in tobacco leaves based on an efficient network structure and multi-scale feature fusion.On one hand, the YOLOv5 (You Only Look Once version 5) network model is modified to extract features in a way that is more conducive to detecting small targets, while reducing the output heads to ensure the lightweight nature of the network. On the other hand, the backbone network and Neck structure are optimized to enable the detection network to obtain richer feature information while maintaining lightweight characteristics. Experimental results demonstrate that the algorithm proposed in this paper not only meets the real-time and lightweight requirements for industrial deployment but also improves accuracy and reduces false positive rates.