Research On The Tea-Weed Localization Method Based On The Improved YOLO Network Model

With the emergent technology industries represented by Artificial Intelligence(AI)flourishing vigorously,in the novel digital technologies extending to traditional industries,Precision Agriculture(PA),supported by automatic and intelligent agricultural machinery(AM),is progressively becoming the significant field of technological innovation and industrial fusion development.Crop protection and weed control are two important tasks of AM.Crop-weed detection is not only the technical challenge in such tasks,but also the core technology for AM crop-weed location application system.The algorithm research of crop-weed localization and classification at present focuses on plant seedlings or rhizomes.However,there is still a lack of tea-weed detection algorithms with high accuracy,good real-time,strong robustness and generalizability,owing to the growth characteristics of tea bushes,the complex light conditions and dense plant distribution in their cultivation environment.To locate tea bushes and green weeds quickly and precisely,based on Computer Vision(CV)and Deep Learning(DL),this research work utilizes an appropriate scheme to collect tea-weed images,and optimizes the network structure and loss function of object detection algorithms,which is beneficial to the technological advancement of the AM crop-weed detection system.The mission objective and performance limitation of the algorithm models,in the training and verifying of DL algorithms,are affected observably by the acquisition method,quality,quantity,and diversity of the samples in Data Sets.Accordingly,the image acquisition method,fit the typical operation viewpoint of the AM in tea plantations,is established by comparison and analysis of shooting directions,the position and angle of a camera,when collecting tea-weed image data.The images,mixed with three varieties of tea bushes and weeds,are acquired in multiple scenes using the method.Then,the tea-weed Data Sets are built by selection and annotation of the images.In contrast to the patterned image collection scheme,this thesis considers the application scenarios of the models and the distribution characteristics of the plants in tea plantations,in order to enhance the robustness,generalization,training quality,and prediction precision of the models.Based on the Data Sets,the research work on AM tea-weed detection algorithms is conducted in two stages.In the early stage,the tea-weed detection algorithm based on the improved YOLOv3 is proposed.Firstly,K-Means++ algorithm is utilized to design the scales of priori anchor boxes.Next,residual network,feature extraction fusion,and Elu activation function are integrated to improve the network structure of the original YOLOv3.Then,generalized intersection-over-union loss and prediction bounding box scale loss parameter are introduced to redesign the loss function,and the relevant hyperparameters are optimized.Finally,the feasibility of the above methods and the detection performance of the improved YOLOv3 model are verified and evaluated by several key performance metrics,such as mean Average Precision(m AP),Precision,Recall,F1-Score,and Frames Per Second(FPS).Following the appearance of YOLOv4 algorithm,the work goes into the next stage.Firstly,the tea-weed detection effect of YOLOv4 is compared with that of the improved YOLOv3.Next,based on YOLOv4,Spatial pyramid block,shared head and complete intersection-over-union are replaced by dense block,un-shared head,and scale-sensitive intersection-over-union loss,respectively,to enhance parameter transfer in the network,solve the conflict problem between classification and regression tasks,optimize the calculation of the boundary frame regression loss,respectively.Then,the improved YOLOv4 model is generated via training on the tea-weed Data Sets.Finally,compared with the improved YOLOv3 model,the improved YOLOv4 model not only has good generalization and robustness but also could achieve higher detection accuracy and faster operation speed.