Tea Disease Detection And Application Based On YOLOv5s-STP In Complex Background

China is a large country of tea production and export.The prevention and control of tea diseases is of great significance to the yield and quality of tea.At present,there are many misjudgments and omissions in identifying tea diseases based on the experience of tea farmers.In recent years,with the development of deep learning,target detection technology has driven the development of crop disease detection,but the existing research on tea disease detection is basically based on single disease detection under a single background.Therefore,it is of great significance to study a multi-leaf tea disease recognition algorithm that can identify multi-leaf tea diseases in a complex background in the natural environment.In order to make up for the vacancy of tea disease detection in the natural environment,based on the target detection technology,this thesis uses the convolutional neural network method to propose a YOLOv5s-STP tea disease detection method for complex background,multi-leaf and small target diseases in the natural environment.The specific work contents and test results are as follows :(1)The tea disease dataset is constructed.In this thesis,a total of 2492 images of 5 diseases of 3 kinds of tea trees are collected,and the dataset is expanded to 7287 by data enhancement methods such as position transformation,color jitter,Auto Contrast,Cutout and Sharpness.Then the dataset is normalized,and then the Lable Img image annotation software is used to complete the image annotation work.Finally,the dataset is randomly divided according to the proportion.(2)A tea disease detection algorithm based on YOLOv5s-STP is proposed.Firstly,the Faster R-CNN,SSD and YOLOv5 s algorithms are used to train the self-built tea disease dataset respectively.The results show that the m AP of YOLOv5 s reached 83.2%,FPS reached 59.6f/s,and the model size is only 14.3M,which is the best performance among the three algorithms.Secondly,to address the problems of unbalanced data samples and poor detection of some diseases,two methods are proposed to optimize the YOLOv5 s algorithm by fusing Swin Transformer Block(STB)into the YOLOv5 s network with C3 module and fusing SPA attention mechanism with residual structure.The YOLOv5s-STP algorithm is constructed to improve the feature extraction ability of the network and reduce the influence of complex background and other factors.The optimized algorithm achieves an isotropic increase in accuracy and recall with almost no effect on the model size,reaching 91.2% and82.4%,which are 6.6% and 0.8% higher than the original YOLOv5 s algorithm,respectively.The value of F1 is increased by 3.5% to 86.6%,m AP@0.5 increased by 3.5% to 86.7%,and FPS reached 56.3f/s.It is confirmed that the YOLOv5s-STP algorithm proposed in this thesis can better identify multi-leaf tea diseases in complex backgrounds on a self-constructed tea disease dataset.(3)A mobile terminal for tea disease detection is designed and developed.The algorithm proposed in this thesis is deployed using the Flask framework and developed using a frontend and back-end separation to realize the functional modules of the tea disease detection system such as detection,history management,and personal center.The mobile terminal is then tested and the results show that the detection platform is available and practical.