SegFormer+: Online Tobacco-Cut Structure Proportion Measurement Method Based on Improved SegFormer

Online tobacco-cut structure measurement can obtain the mass proportion of different types of tobacco-cut in cigarettes in real time, which is of great significance for monitoring the quality of commercial cigarettes. Traditional tobacco-cut structure measurement methods have not only low accuracy but also less efficiency due to complex processes. We first use the Segformer model as the benchmark model, realize the recognition and segmentation of different tobacco-cut (leaf, slice, Dilated) and stems through semantic segmentation, and introduce key improvements including the introduction of snake convolution and the GHM loss function to SegFormer, and a secondary structure proportion revision to further improve the performance of the model, which yields the improved model SegFormer+. In addition, we propose an automated annotation method to produce machine-generated tobacco-cut semantic segmentation datasets, which can obtain training sets without manual annotation so as to save manual annotation costs. Experimental results show that the average relative error of the structure measurement is less than 15%.