TobaccoSeg: A model for determining tobacco components based on semantic segmentation

Online tobacco component determination can obtain the mass ratio of different types of tobacco in cigarettes in real time, which is of great significance for monitoring the quality of finished cigarettes. Traditional tobacco component determination methods have poor timeliness, complex processes, slow speed and low accuracy. We propose the TobaccoSeg model, which uses Segformer as the benchmark model, realizes the recognition and segmentation of different tobacco (leaf, slice, expanded) and stems through semantic segmentation, and introduce snake convolution and GMH loss function to improve the model accuracy and robustness, further improving the segmentation performance of the model. In addition, we propose an automated annotation method for tobacco semantic segmentation datasets, which can obtain training sets without manual annotation, saving manual annotation costs. Experimental results show that the average relative error of the measurement results is less than 10%.