Tobacco Strand Structure Segmentation Method Based on Encoder-Only Mask Transformer Deep Model

In the tobacco production process, in order to address the issues of tobacco strand structure detection, this study investigates the automatic segmentation technology of tobacco strand structure based on deep learning. This study constructed a high-quality and finely annotated tobacco strand image dataset, covering four types of tobacco strands: long, medium, short, and debris. This study is the first to apply the end-to-end visual Transformer architecture EoMT (Encoder-only Mask Transformer) to tobacco strand scenarios, achieving fine segmentation and recognition of various structures in tobacco strand images. Experiments conducted on actual tobacco strand sample batches demonstrate that the proposed model significantly outperforms traditional Convolutional Neural Networks (CNNs) in terms of pixel accuracy (mPA), mean intersection over union (mIoU), and robustness. This research effectively supports the standardization and automated monitoring of tobacco quality by combining advanced technologies and solutions, providing a feasible path and technical basis for intelligent quality control in the tobacco industry.