Summary of Grape Picking Positioning Papers
Positioning of picking points based on improved Yolov5 and natural environments 2023
Identifcation of table grapes in the natural environment based on an improved Yolov5 and localization of picking points
Journal: Precision Agriculture
Abstract: Aiming at the characteristics of table grape picking point positioning in the natural environment, a Yolov5-CBAM-fourth detection based on Yolov5 is proposed Layer decoupled (Yolov5-CFD) network realizes the identification of grapes and stems. At the same time, geometric methods are used to achieve rapid positioning of picking points. First, in order to improve the feature extraction capability of the backbone module, it is improved by integrating the attention mechanism of the convolutional block attention module (CBAM). Secondly, to address the problem of covert detection of a small target, a fourth layer of detection was added to the neck module. In addition, this article draws on the decoupling structure of Yolox to improve the head module of Yolov5 and optimize the classification and regression performance of the network. In addition, geometric methods are used to quickly and accurately locate the picking points of table grapes. To verify the effectiveness of the proposed network model, we used approximately 10,000 grape images for training. The results show that the detection accuracy, recall rate, mAP_0.5 and mAP_0.5:0.95 of the Yolov5s-CFD model are 0.857, 0.804, 0.855 and 0.642 respectively. The detection accuracy, recall rate, mAP_0.5 and mAP_0.5:0.95 of the Yolov5m-CFD model are 0.986, 0.987, 0.993 and 0.910 respectively. Furthermore, the success rate of selected point localization is compared with the corresponding network structure. The results show that compared with the initial Yolov5s, the point selection and positioning success rate of the Yolov5s-CFD model is increased by 11.53%, and the success rate of the Yolov5m-CFD model is increased by 5.84% compared with the initial Yolov5. Although the recognition time of the improved Yolov5 model is increased compared with the initial Yolov5 model, it is still acceptable. It fully meets the requirements for mechanized harvesting of table grapes and provides a theoretical basis for mechanized harvesting of grapes.
Attached picture:
First, Yolov5 v6.0 is improved to train and detect table grapes and grape stems. Then, the information of the grape stems was extracted separately. Finally, the picking points of table grapes were obtained using geometric methods. An overview of the proposed method is shown above.
The improvement is mainly in the Backbone, Head and Neck parts: In order to avoid the loss of effective feature information caused by multiple convolutions, a CBAM module is added at the end of the backbone module to improve the Head model and perform classification and regression tasks respectively; in the Neck module, a A detection layer for small object recognition.
The grape and fruit stem feature maps of feature pyramid FPN and pixel aggregation network PAN are output as the input of the detection layer, which is the 21st, 24th, 27th, and 30th layers of the model.
After input, Conv2D_BN_SiLU processes the input feature map, which consists of a convolution layer, a BN layer and a SiLU activation function.
The pixel ratio of long-distance grape stems is low. Grape stems are extracted through K-means clustering algorithm, OTSU algorithm and color adjustment method. The
aspect ratio of the stem image is used as the basis for stem classification to achieve picking positioning:
1. When the height/width is greater than or equal to 3, the fruit stem is considered to be vertical and unbent, and the midpoint of the fruit stem image is used as the picking point. 2. When the height/width is less than 3, the fruit stem is considered to be
curved. The bending point is located in the upper part of the image and the color is darker than the background. Use the point directly above the center of the image as the picking point
The following are ablation experiments:
Different experiments on yolov5
Results of different network models
Yolov5s, Yolov5m, Yolov5l, Yolov5x, Yolov3, Yolov3-SPP, Yolov7,
Yolov5s-D (Yolov5s-Decoupled), Yolov5s-F (Yolov5s-Fourth Detection Layer), Yolov5s
FD (Yolov5s-Fourth Detection Layer-Decoupled), Yolov5s-CFD (Yolov5s-CBAM-Fourth
Detection Layer-Decoupled), Yolov5m-CFD (Yolov5s-CBAM-Fourth Detection Layer
Decoupled).进行对比