Tea bud pose estimation and grading detection network based on improved YOLOv7

Yuchen Yao^1,2Gui Zhiyong²Haoyang Liu^2,3Zidong Yang¹Lijian Yao¹Kai Li¹Zhenchuan Lin⁴Yihu Mao²Zhijun Jia^1,2Yang Li^2,3*Rong Ma^1*

• ¹Zhejiang A and F University, Hangzhou, China
• ²Tea Research Institute, Chinese Academy of Agricultural Sciences, Hangzhou, China
• ³National Key Laboratory for Tea Plant Germplasm Innovation and Resource Utilization, Hangzhou, China
• ⁴FUJIAN PIN PIN XIANG TEA INDUSTRY CO., LTD, Ningde, China

Intelligent recognition and rapid grading of tea buds are significant for the development of tea-picking machinery. However, due to the complex background of tea plantations and the inconsistent growth of tea buds, traditional recognition algorithms have only ad-dressed the issue of identifying picking points, without fully considering the pose and grade of buds. As a result, the picking end-effector cannot adaptively harvest tea buds based on their pose and grade, thereby limiting the success rate and efficiency of picking. To address this issue, we propose YOLO-PC (You Only Look Once-Pose estimation and Classification), a deep neural network for tea bud pose estimation and classification. This model enhances shape feature extraction using a dynamic snake convolution (DSConv) module and improves spatial pooling with an ELA-SPPCSPC attention mechanism. Replacing CIoU with EIoU accelerates regression and boosts localization accuracy. Experiments show 91.5% accuracy for one-bud-one-leaf detection and 93.2% for one-bud-two-leaf detection, with an average keypoint detection accuracy (Pose_mAP) of 89.7%. The Normalized Mean Error (NME) achieved 0.047. The improved model compared to YOLOv7-pose increases mean average precision by 7.26% and pose accuracy by 9.65%, while reducing parameters by 14.99 M. Ablation assays confirm its superior performance in tea bud detection, providing practical support for intelligent harvesting.