Deep Learning Based Tea Picking Point Location

The history of tea drinking in China can be traced back to as early as the Qin and Han dynasties,and after 2000 years of development,tea drinking has become a habit for most Chinese people.In tea production,tea bud picking is a particularly critical aspect,because the quantity and quality of tea buds directly affects the scale of the entire tea industry.At present,tea picking methods are mainly manual and mechanical,but the efficiency of manual tea picking is low,and the integrity rate of mechanical tea picking buds is also low.Therefore,there is a need to improve the existing tea harvesting methods or to develop new ones.One of the most effective ways to achieve mechanical intelligent tea harvesting is to enable the computer to "recognize" the tea shoots and stems,which can ensure both high efficiency and high bud integrity rate.This thesis constructs a tender tea stalk recognition model based on deep learning algorithm using TEA STEM dataset,and the main research content is as follows:(1)Multiple shoot tea stalk recognition models were constructed for experimental comparison to obtain the best performance of YOLOv7 algorithm.In this thesis,we constructed the shoot tea stalk recognition models based on PSPnet,Unet,Deeplab V3+ and YOLOv7 algorithms on TEA STEM dataset.After the experimental comparison and evaluation analysis,it was found that the YOLOv7 algorithm had the best performance with90.89% m AP.(2)Replacing YOLOv7 convolutional network with Ghost Conv and adding Sim AM attention mechanism.In this round,Ghost Conv is used to lighten YOLOv7 and add ECA,Sim AM and CBAM attention mechanisms respectively,and the final experimental results show that YOLOv7+Ghost Conv+Sim AM has the best performance and its m AP reaches92.13% after hyperparameter tuning.(3)Constructing a 3D spatial shoot picking point recognition model.The YOLOv7 algorithm combined with the binocular camera was used to construct a 3D shoot recognition model,which was combined with the shoot stalk recognition model of(2)to determine the3 D coordinates of the shoot picking point with an accuracy of 84.4%.(4)Develop a web-based shoot tea stalk recognition system.It integrates tea shoot dataset processing,image annotation,model training,target recognition and model evaluation functions.This thesis constructs a tender tea stalk recognition model and calculates the depth information of tea stalks in three-dimensional space,which proves the feasibility of mechanical intelligent tea picking and provides some theoretical support for the subsequent application of mechanical intelligent tea picking in practice.