Abstract: To reduce the time cost of repeatedly constructing data samples of tea buds from different seasons and regions, and to improve the accuracy and generalization ability of deep learning models in identifying different tea buds, this paper proposed a small-sample tea bud recognition model based on deep transfer learning, named QY-Yolov7-Tea. By incorporating transfer learning techniques into deep learning, a source domain bud detection model based on Yolov7 was used to obtain pre-trained weights. The model's backbone, neck and detection head are then fine-tuned and frozen, with experiments conducted on different target samples based on the pre-trained weights. The final result is a transfer learning model for tea buds. Experimental results show that, compared with the traditional Yolov7 bud detection model, the deep transfer learning model improves the mean average precision for recognizing tea buds from different seasons and regions by 8.8 percentage points and 15.4 percentage points, respectively, significantly enhancing the model's robustness and recognition capability, and effectively addressing the issue of insufficient training samples.

Key words: tea bud, deep learning, transfer learning, model fine-tuning