Segmentation of tobacco shred point cloud and 3-D measurement based on improved PointNet++ network with DTC algorithm

The three dimensions of the tobacco silk components (cut stem, tobacco silk, reconstituted tobacco shred, and expanded tobacco silk) of cigarettes directly affect cigarette combustibility; by accurately measuring the dimensions of different tobacco silks in cigarettes, it is possible to optimize combustibility and reduce the production of harmful substances. Identifying the components of tobacco shred in cigarettes is a prerequisite for three-dimensional measurement. The two-dimensional image method can identify the tobacco shred and measure its two-dimensional characteristics but cannot determine its thickness. This study therefore focuses on the identification of the tobacco shred and measuring it in three dimensions.

The point cloud data of the upper and lower surfaces of tobacco shred are segmented using the improved three-dimensional point cloud segmentation model based on the PointNet++ network. This model combines the weighted cross-entropy loss function to enhance the classification effect, the cosine annealing algorithm to optimize the training process, and the improved k-nearest neighbors multi-scale grouping method to enhance the model's ability to segment the point cloud with complex morphology. Meanwhile, this study also proposes a dimension transformation calculation method for calculating the three dimensions of tobacco shred.

The experimental results show that the precision and recall of the improved segmentation model increased from 84.27% and 83.63% to 95.13% and 97.68%, respectively; the relative errors of the length and width of tobacco shred were less than 5% and 7%, and the relative error of the standard gauge block thickness measurement reached 1.12%.

This study also provides a new idea for implementing three-dimensional measurements of other flexible materials.