Abstract
Tea winnowing is a key process in tea processing. At present, tea winnowing parameters are adjusted by manual observation of tea leaves. This results in the uncertainty of winnowing quality. In this work, we propose a new tea winnowing method based on deep learning for the characteristics of white tea. Firstly, the YOLOv11 model is improved by introducing ACmix and EUCB. The recognition accuracy of the improved YOLO-AE model is improved by 2.1%, and the detection time is shortened by 40%, which significantly improves the detection performance and shortens the inference time. The region segmentation and convolution neural network algorithm are used to distinguish the proportion parameters of each grade in tea in real time, and the accurate wind selection parameters are obtained by combining the winnowing theory. The recognition accuracy of the verification set of the recognition model attains 94%. The MAP (0.5:0.95) is 0.93. A test on the tea winnowing parameter control test bench reveals that the identification accuracy of tea materials with different proportions is consistent. Additionally, the difference between the two batches of high-quality white tea is less than 3%. The winnowing scheme proposed in this study can provide the basic theory and technical support for the design of tea precision winnowing equipment.