Abstract: Accurate identification of tea shoots in a complex background is one of the key technologies to realize the intelligent picking of high-end tea. In order to realize the mechanized and precise picking of high-end tea, this paper designs a visual-based tea picking prototype, which converts the moving coordinate problem at the end of the manipulator into the corner problem of three motors of the static platform according to the path planning of the spider manipulator picking tea. The YOLOv3 algorithm is improved, the EfficientNet network is used instead of the DarkNet-53 network for feature extraction, and the objective function GIOU is used to optimize the loss function. The experimental results show that the improved YOLOv3 algorithm has an accuracy rate of 86.53% in tea bud recognition, and the average recognition time for a single image is 53 ms. Compared with the traditional YOLOv3 algorithm, the performance has been significantly improved, which can achieve the expected goal and meet the needs of machine picking.