烟叶烘烤阶段不同YOLO算法模型的实时判别性能比较OA

[Objective]This study aims to compare and explore the applicability of YOLO algorithm models in the discrimination of tobacco leaf curing stages and identify the optimal version.[Method]Four versions of YOLO,including YOLOv3,YOLOv5s,YOLOv8s,and YOLOv11s,were selected to construct models respectively.These models were compared and evaluated based on three aspects:dis-crimination performance,complexity,and real-time detection ability.[Result]The YOLOv8s model performed better overall than the other three versions,with an precision of 95.0%,an average accu-racy value of 97.1%,and an F1 score of 93.0%,all of which were the best among the models.The model's recall rate was 91.1%,which was slightly lower than the YOLOv5s model's 93.3%.Addition-ally,the YOLOv8s algorithm demonstrated the fastest detection speed,with a detection time of only 5.5 ms per image.Notably,the YOLOv8s algorithm model showed outstanding performance in classi-fying stages with subtle changes in tobacco leaf condition,such as the stem drying stage.In terms of lightweight design,YOLOv5s performed the best,with a parameter count of 7.04×106,a computa-tional load of 15.8 G FLOPs,and a model size of 13.7 MB.Under comparable power consumption levels across all four models,the YOLOv5s model required 1 081 MB of runtime memory,only slightly higher than the 916 MB required by the YOLOv8s model.[Conclusion]Through comparative analysis,significant differences were observed in the performance of the four YOLO algorithm versions in dis-criminating tobacco leaf curing stages,with YOLOv5s and YOLOv8s performing better.YOLOv8s exhibited the best overall performance,while YOLOv5s had lower computational load and model size,making it more suitable for low-cost embedded devices requiring lightweight models.Therefore,YOLO algorithm models are suitable for real-time discrimination of tobacco leaf curing stages,and the version selection should be based on specific needs.