Study on Improved YOLOv8 Real-time Dynamic Visual Detection Algorithm for Tobacco Clusters Based on Mosaic Data Enhancement and Difficult Sample Mining

Aiming at the product quality problems caused by the agglomeration phenomenon during the production of tobacco in cigarette factories, this paper proposes a real-time dynamic visual inspection algorithm based on YOLOv8. By combining Mosaic data enhancement and difficult sample mining methods, it solves the challenges of small number of agglomerated samples, high real-time requirements for detection, and difficulties in detecting agglomerates of small sizes. The experimental results show that the improved YOLOv8 algorithm achieves 95.6% detection accuracy and 94.0% detection rate on the test set, which is significantly better than the traditional Faster R-CNN (89.6% accuracy, 92.9% detection rate) and the original YOLOv8 (93.1% accuracy, 91.2% detection rate). After the actual production line deployment, the number of false alarms of the system is less than 10 per month, and the success rate of manual simulation test is as high as 99.4%. This study provides an efficient automated solution for the quality control of tobacco filament agglomeration, which has significant application value.