Research On Accurate Localization Of Tea Tree Bud Picking Point Under Natural Scenes Based On Deep Learning

As a traditional healthy drink in China,tea is deeply loved by people,and high quality tea is increasingly sought after,the demand for premium tea is growing.The traditional manual picking method is time-consuming and inefficient,while the fully mechanized picking method is "one-size-fits-all",which is easy to harm the quality of tea leaves.With the growing application of artificial intelligence technology in the field of agricultural picking,tea picking methods based on deep learning are also receiving more and more attention.Different from the picking of other crops,premium tea picking faces the problem of difficult detection caused by the small target of buds and leaves themselves,varying morphology and density,and mutual obscuration between buds and leaves,as well as the problem of small picking points that are difficult to locate precisely.This thesis takes highquality tea in natural environment as the research object,and focuses on the detection of tea tree buds in picking and the precise location of picking points.The main research contents are as follows:(1)Rapid detection of tea tree buds in the non-structural environment of tea plantations.In order to achieve the task of accurate detection of tea tree buds under natural lighting in tea plantations,images of tea tree buds under natural environment are collected using depth cameras and digital cameras,and image data are augmented to construct tea tree buds detection and segmentation datasets.Based on the YOLOv5 s target detection framework for bud leaf detection,the ACmix module with convolution and self-attention fusion is introduced to increase the feature extraction capability of the backbone network for bud leaf detection,and the YOLOv5s-ACmix bud leaf detection model is constructed.Experiments show that the improved tea tree buds detection models AP and Recall are84.04% and 67.91%,respectively,and the detection frame rate is 21.8 FPS.AP is improved by 0.74% compared with the original model,respectively,and the network maintains a better detection speed while improving the detection effect.(2)Two-dimensional localization of tea tree buds picking points based on the multitasking model.In order to complete the accurate positioning of picking points and realize the determination of two-dimensional picking points through the intersection of tea tree buds detection frame and tea stems,a simple segmentation head is introduced in the improved YOLOv5s-ACmix model to output the tea stem region and share the backbone network to construct a multi-task model based on the improved YOLOv5 s for tea tree buds detection and tea stem segmentation and determine the two-dimensional plane coordinates of picking points.Experiments show that the model detects AP up to 84.63%,segmentation mIoU up to 52.04% and frame rate 16.8FPS,which can meet the demand of fast detection and accurate localization of famous tea picking in the field.(3)Three-dimensional localization of tea tree buds picking point based on depth camera.The RGB images acquired by the depth camera are input into the constructed tea tree buds detection and tea stem segmentation model to determine the two-dimensional plane coordinates of the picking point,and acquire the depth images aligned with the RGB images,and calculate the coordinates of the three-dimensional picking point under the camera coordinate system through the depth data information of the depth images to realize the three-dimensional positioning of the picking point,providing a technical solution for the intelligent picking of tea tree buds for the accurate positioning of the picking point.