Objective:In order to realize the intelligent grasping of tobacco stem in tobacco grading process, prevent the manipulator in the intelligent tobacco grading system from damaging the leaf surface during grasping tobacco leaves, and reduce the manual operation in the production of intelligent tobacco grading equipment.Methods:An automatic tobacco stem identification and location model based on improved YOLOv3 convolution neural network was proposed for the identification and classification of single tobacco leaf and the storage of corresponding single tobacco leaf in tobacco grading system. The model changed the structure of the unit module and introduced the attention mechanism module based on the original YOLOv3 model, which optimized the model parameters and used swish activation function to realize the target location and recognition of all the information of tobacco leaf images, and then the tobacco stem target detection model was constructed.Results:The results showed that the loss of improved YOLOv3 model could converge faster, with its mAP increased from 90.46% to 97.48% and its accuracy increased from 95.33% to 97.35%; its regression rate increased from 84.65% to 95.65%, which laid the foundation for the automatic classification of tobacco leaves.Conclusion:Compared with YOLOv3, Faster-rcnn, YOLOv4, Efficientdet algorithm, the proposed algorithm is lighter and more effective. It can reduce the hardware configuration requirements of tobacco stem test platform, improve the economic benefits of tobacco classification system, and provide accurate location information for tobacco feeding and storehouse separation in tobacco classification system.