Abstract: In order to improve the efficiency and accuracy of identification of tobacco stems (pure stem, stem head and lamina-attached stem) during green leaf threshing, an on-line classification and identification model for tobacco stems was established based on deep learning. First, a digital image processing method was used to extract the target from the collected tobacco stem images and create a tobacco stem data set. Second, the original YOLOv3 model was revised according to the characteristics of tobacco stem images to configure a new network structure. Then, the revised YOLOv3 model was trained with the created data set to establish a deep learning-based classification and identification model for tobacco stems. Finally, the established model was loaded into an on-line tobacco stem classification and identification system to verify its performance. The results showed that the established model performed well on the test set, the identification accuracy of the established model reached 95.01%, it was 5.97 percentage points higher than the original YOLOv3 model, and the recall rate increased by 4.76 percentage points, both were better than those of SSD and Mask R-CNN models. Featuring a strong anti-interference ability under different complex scenarios, the established model is effective to identify the locations and categories of tobacco stems and competent for the rapid classification and identification of tobacco stems. This method provides a technical support for promoting the classification efficiency and identification accuracy of tobacco stems.