一种基于改进YOLOv4算法的茶树芽叶采摘点识别及定位方法

In view of the difficulty in identifying and locating the picking point of tea bud leaves in complex environment with the agricultural tea picking robot quickly and accurately, the improved YOLOv4-Dense algorithm and the OpenCV image processing method are used to study the positioning of tea bud leaf picking points. Firstly, based on the YOLOv4 algorithm, the ResNet unit in the CSPDarknet53 trunk feature extraction network is replaced with the DenseNet unit, then the improved algorithm model was used to detect the tea bud leaves of the collected tea plant dataset. Secondly, the color channel conversion of RGB to HSV is performed to obtain the outline of the bud leaves by using OpenCV, then the location of picking points is located by OpenCV morphological processing. Finally, conduct comparative experiments on picking point positioning methods and compare with the positioning results of the moment function method and the smallest external rectangular center point method. The results show that: 1) The precision of the improved YOLOv4-Dense algorithm in the detection of tea bud leaves was 91.83%, the recall rate was 68.84%, the AP was 86.55%, and the F1 score was 0.79; Compared with the YOLOv4 model, the precision, recall, AP value, and F1 score improved by 2.21%, 2.00%; 2.05%, and 0.02; Compared with the YOLO v3 model they improved by 5.56%, 15.26%; 9.13% and 0.13; 2) For tea bud leaves under natural conditions, the accuracy of locating the picking point using OpenCV image processing method was 80.8%, and the recall rate was 83.2%, compared with the moment function method and the minimum external rectangular center point method, it is increased by 3.5%, 7.1%; 1.4%, 6.1%, indicating that this method has certain reference significance for the accurate identification and positioning of tea bud leaves picking points.