TeasFormer: An Improved Tea Classification Algorithm Based on Pyramid Vision Transformer

This study presents an improved tea classification model algorithm, named TeasFormer, developed from the Pyramid Vision Transformer(PVT). This algorithm integrates the DCFNN module, MSFF module, and Residual module, and employs a transfer learning strategy to transfer knowledge learned from the Mini-ImageNet dataset to target data. This enhances the model's capability to recognize different types of tea. Experimental results indicate that the TeasFormer, model achieves a accuracy of 94.8%, precision of 94.6%, a recall of 94.1 %, and a F1 score of 94.3% in identifying ten distinct types of tea. These performance metrics exceed those of existing models such as ResNet34, Vision-Transformer, Swin-Transformer, MobieNet, and EfficientNet. This research provides an effective technical foundation for tea identification and authentication, promising broad applications within the tea industry.