An end-to-end lightweight model for grape and picking point simultaneous detection

Grape clusters and their picking point detection (GCPPD) are pivotal in the visual tasks of automatic grape harvesting. In recent years, much progress has been made in increasing the accuracy of GCPPD based on deep learning models. However, GCPPD still has many problems. First, it is inevitable that grape cluster detection requires complex models with many parameters. Second, the prior work on picking point detection can be summarised as the image processing methods using predefined hand-crafted features. This leads to a lack of robustness in the proposed algorithms. To address this, a scheme for the simultaneous detection of grape clusters and their picking points is explored. Due to the superiority of simultaneous detection, the model is constructed as an end-to-end network. Thus, a lightweight end-to-end model called YOLO-GP (YOLO-Grape and Picking points) is proposed. Specifically, YOLO-GP utilises a ghost bottleneck to reduce model parameters. Additionally, this model adds the prediction of picking points using the novel idea, that the picking point follows the bounding box. The Grape-PP (Grape-Picking Point) dataset for model training is constructed, which contains 360 grape images with 4517 grape cluster bounding boxes and picking points. The experiments show that the mean Average Precision (mAP) of grape cluster detection by YOLO-GP is 93.27% with a decrease in the number of weight parameters by at least 10%. The distance error of picking point detection is less than 40 pixels. In summary, YOLO-GP achieves the simultaneous detection of grape clusters and their picking points, and its performance is comparable to that of baseline models.