Few-Shot Learning for Small Impurities in Tobacco Stems With Improved YOLOv7

Abstract

With the increase of public concern about health and smoking, the authorities have gradually tightened the control of tar content in cigarettes, making reconstituted tobacco a growing concern for tobacco companies. Tobacco stems are used as the main raw material for reconstituted tobacco, but they contain a large number of small broken impurities mainly from cigarette butts, which are difficult to remove efficiently by air selection and manual methods. Detection schemes for cigarette butt impurities based on computer vision and deep learning are still difficult. The scarcity of images containing foreign impurities in cigarette butts and the small size of impurities limit the efficient application of deep learning algorithms. In view of the small impurities’ characteristics, this paper optimizes the model structure of the YOLOv7 algorithm, and only retains the two detection head structures with high feature resolution, which reduces the model parameters by 29.68%. Using online data augmentation and transfer learning, the difficulty of small sample datasets is overcome. After the CutMix, Mosaic, Affine transformation, Copy-paste data augmentation in this paper, the model precision is increased by 6.95%, and the recall rate is increased by 10.51%. Detection FPS has been increased from 99 FPS to 111 FPS. Precision and recall rate reached 97.21% and 92.11%. Compared with YOLOv4_csp, the precision is in-creased by 11.58%, and the recall rate is increased by 0.48%. It shows that the improved YOLOv7xs model has the potential for wide application in small target recognition. At the same time, it has shown the potential to avoid the harm of toxic substances produced by cigarette impurities in the combustion process and promotes the application of computer vision and deep learning in industrial production.

Keywords

• Impurities detection
• deep learning
• computer vision
• tobacco
• data augmentation
• YOLO-v7