Research On Identification And Picking Location Of Famous Tea Bud Based On Computer Vision

As one of the local specialties of Anhui Province,famous tea is popular with consumers for its fat sprouts,fresh taste and elegant aroma.However,the current mainstream "one size fit all" mechanical tea picking method does not have selective picking ability,and the integrity of tea after collection is poor,which cannot meet the requirements of producing high-quality famous tea,resulting in a sharp decline in the economic income of tea farmers.In view of the above problems,this topic is based on computer vision technology to identify and locate the famous tea buds in the natural environment,so as to provide technical support for the subsequent research and development of automatic and intelligent picking equipment.The main research contents are as follows:(1)Based on taking pictures of different kinds of tea trees in the natural environment,the training and testing data sets of the model are constructed.According to different times and weather conditions,the pictures of famous tea are obtained by binocular cameras,mobile phones and other devices,and the bud positions in the pictures are marked by Labelimg software.On this basis,the data set is expanded by rotation,translation,adding noise and other methods to increase the amount and diversity of training data of the model,and improve the generalization ability and robustness of the model.(2)An improved YOLOv7 algorithm is proposed to improve the recognition accuracy of famous and excellent tea buds in complex environments.Firstly,the original YOLOv7 backbone network was replaced with Mobile Netv3.Then,GAM global attention mechanism and cosine annealing learning strategy were introduced to improve the accuracy and robustness of the model.Finally,the constructed data set is used for comparative verification,and the results show that: The m AP and FPS indexes of the improved YOLOv7 are 98.07% and 83,respectively,which are 1.34% higher than the m AP value of the original YOLOv7,the FPS is increased by 9.21%,and the model size is reduced by 70.83%,which are far better than the Fast R-CNN and SSD models.The proposed improved algorithm can accurately identify the famous and excellent tea buds,has real-time detection ability,and meets the requirements of intelligent picking to a certain extent.(3)Based on the accurate identification of the bud subject,this thesis analyzed and compared the characteristics of the bud subject under the RGB,HSV and Ycrcb color models,and found that the color image of the famous tea bud image and the background image were significantly different under the RGB color space.To this end,the G-B channel components of the bud body are extracted and the feature maps are generated.Then,Otsu method algorithm is used to segment the main bud in the feature map,and the twodimensional coordinates of the picking point are obtained by morphological processing.(4)Based on binocular stereo vision and improved stereo matching algorithm,the picking point location of famous tea bud is realized.Based on the imaging principle of binocular camera,the internal and external parameters of ZED binocular camera were obtained by using Zhang Zhengyou calibration method.On this basis,the SGBM algorithm was used to perform pixel stereo matching on the famous tea bud images,and then the disparity was calculated and the depth map was obtained.Then,the two-dimensional coordinates of the picking point and the depth image were transformed into the required three-dimensional coordinates by using the triangle ranging principle.Finally,the experiment results show that the relative error between the distance measured by ZED binocular camera and the actual distance is less than 2%,which provides a more accurate data basis for the research and development of robotic arm positioning picking platform.