Research And System Development Of Tea Bud Recognition And Location Based On Computer Vision

Tea has high economic value.During tea production,the tea picking process is crucial as the quality of the picked tea leaves directly affects the quality of the final tea product.Rapid identification and localization of tea buds are important for achieving intelligent tea picking.In this study,based on computer vision,this paper conducted research on tea bud segmentation and recognition,2D and 3D localization,and developed a tea bud recognition and localization system to lay the foundation for achieving intelligent tea picking.The main research content and conclusions of the thesis are as follows:(1)Establishment of a segmentation recognition model for tea buds.A new method for tea bud segmentation and recognition was proposed,which uses an improved lightweight Deep Lab V3+ model with Mobile Net V2 network as a feature extraction network,which incorporates multi-attention mechanisms,and uses Focal Loss to correct class imbalance.Through ablative experiments and comparison experiments with segmentation models,the model achieved an m Io U of 88.02%,m PA of 93.17%,and FPS of 59.01.The model improved segmentation accuracy and speed,distinguishing the tea bud from the background in the tea bud image and effectively eliminating background interference.(2)Establishment of a 2D localization model for tea buds.Based on the YOLOv7 algorithm,a tea bud 2D localization model was constructed.Before model training,a tea bud segmentation dataset was constructed using the tea bud segmentation model to remove background interference.YOLOv7,YOLOv3,YOLOv4,YOLOv5,and YOLOX were used to train the tea bud segmentation data samples.The experimental results showed that the YOLOv7 model performed the best,with an m AP of 94%,meeting the requirements of tea bud localization.Based on the YOLOv7 model,the 2D coordinates of the tea bud picking points were calculated.(3)Based on the RAFT algorithm,a tea bud 3D localization model was constructed.To calculate the 3D coordinates of the tea bud,the tea bud segmentation image was obtained through the segmentation model,the 2D pixel coordinates of the tea bud were predicted through the tea bud 2D localization model,and the stereo images of the tea bud were subjected to stereo correction,the depth information of the tea bud stereo images was calculated through the RAFT stereo matching algorithm,and the 3D world coordinates of the tea bud were obtained through coordinate transformation.The feasibility of the 3D localization approach was demonstrated by comparing the RAFT algorithm and the SGBM stereo matching algorithm,and by using the built binocular system to calculate the 3D coordinates of the tea bud picking points.(4)Development of a tea bud recognition and localization system.Based on the actual requirements of intelligent tea bud picking in the future,a tea bud recognition and localization system based on Pyqt5 was designed,which achieved the tea bud segmentation and recognition,2D and 3D localization functions.The accuracy and speed of the system’s recognition and localization were tested based on the tea bud stereo images obtained through the binocular image collection system.The results showed that the tea bud segmentation accuracy was 86.60%,the final tea bud localization accuracy was 84.85%,and the tea bud recognition speed was 32.5ms per tea bud,the 2D localization speed was 39.1ms per tea bud,and the 3D localization speed was 129.2ms per tea bud,effectively satisfying the requirements of tea bud recognition and localization.