Research Of Tobacco Scatter Degree Detection Based On YOLOX

In the process of tobacco leaf grading,it is necessary to manually lay the tobacco leaves loosely,which may lead to the uneven spread of each tobacco leaf during the process of tobacco leaf scattering,and a large number of leaf overlap phenomena.However,the degree of tobacco leaf scattering is too low,which will obscure the key information about tobacco leaf leaves,and affect the quality and stability of tobacco leaf sorting.In order to solve this problem,this paper uses the object detection method in deep learning to study the tobacco leaf scattering situation,including the following points:1.In order to fit the real environment of tobacco leaf scattering detection and obtain tobacco leaf data under the real production environment,this paper simulates the real environment of flue-cured tobacco leaf sorting on the production line,transmits tobacco leaves through the conveyor belt,and builds an image acquisition system to shoot the moving tobacco leaves on the conveyor belt in real-time.In addition,the dataset studied in this paper is obtained by using Open CV image processing technology for the photographed tobacco leaf dataset,and then mosaic and mixup data enhancement is performed.2.Aiming at the problem of large workload and high cost of manual image data labelling,this paper proposes a semi-automatic image data labelling method based on the YOLOX object detection algorithm,which uses model prediction information to realize data labelling and reduces the workload of manual data annotation.3.According to the problem requirements of the current flue-cured tobacco leaf scattering situation,the existing deep learning target detection algorithm is analyzed in depth,and the loose rate is defined to represent the degree of tobacco leaf loosening based on the actual requirements of this paper,and the definition and calculation formula of the scattering rate are given,as well as the process of loosening detection.4.Based on the YOLOX object detection model,this paper constructs a tobacco leaf scatter detection model,and optimizes the model in many aspects,improves the model detection effect,and proves the effectiveness of the improved model through ablation experiments.The average accuracy of the improved model m AP0.5 is86.0%,m AP0.95 is 47.7%,and m AP0.5 and m AP0.95 increased by 2.8 and 3.9percentage points,respectively.The proposed algorithm is compared with the existing mainstream target detection algorithms,and the proposed algorithm has certain advantages in terms of the calculation amount,parameter quantity and average accuracy.5.Based on the tobacco leaf scatter detection model built in this paper,this paper combines the data acquisition system to build an online tobacco leaf scatter detection system for real-time detection of tobacco leaves on the conveyor belt on the production line.