| In contemporary society,tea has become a necessity of people’s lives.With the increasing demand for tea,tea yield has undoubtedly become the object of concern.Therefore,the count of tea buds has become an important part of regional output estimation and tea phenotype research in our country and the world.In the past,the estimation of tea production basically relied on the relevant agricultural workers to go into the wild tea garden to manually count,but this counting method is inefficient,because it is not only time-consuming and laborintensive,but also has strong subjectivity,which is not conducive to the estimation and analysis of tea production.The traditional method of identifying and counting tea buds mainly uses artificial extraction of feature information,such as determining the tea bud identification model according to the color properties and the shape properties of tea buds.Although this method can roughly distinguish the tea buds,the recognition performance of the model will be greatly reduced in a more complex environment.Confronted with these problems,a majority of researchers have paid great attention to the intelligent identification and counting of tea buds.An effective tea counting method requires a sound solution to the complex background environment(lighting,debris,etc.),target adhesion and occlusion in the tea image.In deep learning,Convolutional Neural Networks have powerful feature extraction capabilities,which greatly reduce the influence of environmental factors and the deficiency of manual feature extraction,providing a reference for the intelligent identification and counting of tea buds.This research takes Chongqing Yongchuan Xiuya tea as the research object,using the CNN model in deep learning to identify and count the buds of the collected tea pictures.The main research contents are as follows:(1)A method for counting tea sprouts based on the principle of regression density map per unit area is proposed.In the experiment,the different tea data in the two periods were processed into two data sets with different sizes,including the larger area size of one-quarter of the unit area and the smaller area size of the two tea sprout data sets.The density image of the tea picture was obtained by training the model,and then the density image was returned to obtain the number of tea buds in the picture.The experiment compared the model’s counting performance on tea data of different periods and different sizes,and also compared the performance of the multi-column Convolutional Neural Network MCNN.The results show that the CSRNet model has better counting ability,the counting accuracy and the robust performance of the model are better than MCNN,in which the counting accuracy of tea buds is about 90%;and in a quarter of the unit area of the tea buds counting effect of the model is better,its counting accuracy exceeding 90%.This provides technical means for the estimation of tea production in tea gardens.(2)A tea bud identification and counting method based on the idea of target detection in unit area is proposed,and a small-scale tea bud local counting idea is designed according to the actual situation of the tea data set.The limitation of the model for counting shoots in a larger range is solved.The experiment compared the performance of the two mainstream target detection model algorithms Faster RCNN and Yolo-V3 and the performance of different tea data sets.The results suggest that the bud identification and counting method based on the Faster RCNN model has better overall performance,and its model is better than the Yolo-V3 model in terms of robustness and counting accuracy for tea bud identification.Among them,both recognition accuracy and counting accuracy of the Faster RCNN model for buds are as high as about 95%.In addition,the model can accurately obtain the specific location information of the tea buds in the picture,which lays a solid foundation for the intelligent picking of tea gardens and the counting and estimation of tea production in the future. |
PDF Full Text Request