Research On Tea Bud Detection Method Based On Deep Learning

Tea buds are deeply appreciated for their excellent quality and precision processing.However,the picking method of tea buds is still in the stage of manual picking,which leads to the problems of high labor intensity,high cost and low efficiency and cannot meet the market demand.With the wide application of modern science and technology in the agricultural field,automatic picking equipment based on deep learning is used for intelligent picking of tea buds.The most challenging and key problem in the process of equipment development is how to accurately identify and locate tea buds and embed the model into the mobile terminal to realize real-time detection.Aiming at the problems that the existing tea bud detection algorithms have low detection accuracy,poor generalization performance and can not realize real-time detection,this paper took Longjing 108 as the research object(Tea Research Institute of Chinese Academy of Agricultural Sciences)and put forward an intelligent identification algorithm of tea buds based on deep learning.The research contents mainly include the following aspects:(1)Construct datasets of tea buds growing in natural environment.Aiming at the tea buds collected in different periods,different illumination and different angles,Label Img software was used to label the tea bud images,and various data enhancement technologies were applied to expand the scale of the dataset,and Python programs were written to convert the dataset into a format that can be recognized by neural networks.(2)Select the baseline network and analyze the shortcomings of the baseline network.The current mainstream object detection networks were experimented on selfmade datasets,and compared the experimental results.Considering the detection accuracy and field detection requirements,the reasons for choosing YOLOv3 as the baseline network were analyzed,and analyzed the shortcomings in detecting tea buds.(3)Build YOLO-Tea deep learning network framework based on YOLOv3.Aiming at the problem of missed detection and false detection of tea buds caused by insufficient feature extraction ability of the original YOLOv3 network,the main feature extraction network,feature fusion network and prediction head of the original YOLOv3 network were improved respectively.Aiming at the problem that the bounding box of the original YOLOv3 network was not located accurately,the anchor box,bounding box regression loss function and candidate box filtering algorithm of the original YOLOv3 network were improved respectively.Through different data enhancement algorithms,the Token Mix data enhancement method was innovatively applied to YOLOv3 network,which enhanced the robustness and generalization performance of the model.The experimental results showed that compared with the original network,the improved YOLO-Tea network effectively improved the detection performance of tea buds under complex background.(4)Build a lightweight network YOLO-Ghost based on YOLO-Tea network.Aiming at the problems of YOLO-Tea network in embedded devices,the original network was improved and designed a lightweight model,the specific improvement points include: 1)the main feature extraction network Dark Net53 was replaced by Dark Net19 to improve the model training speed and reduce the memory occupation;2)a lightweight multi-scale convolutional block module(MCBAM)was introduced into the network,without increasing the parameters of the model,the information that was beneficial to the classification of tea buds was gathered and the irrelevant information was suppressed;3)Ghost module was used to replace the ordinary convolution block of the original network,and generated the multi-feature map by cheap operation,which reduced the parameters of the model and improved the accuracy of the model.The experimental results showed that compared with the original YOLO-Tea network,the improved YOLO-Ghost network achieved real-time detection while maintaining high accuracy.