Enhancing small-object detection in vineyards: a YOLOv8-based approach for grape maturity classification

ABSTRACT: ‘Kyoho’ grapes are renowned worldwide for their excellent fresh food quality. Determining the optimal time to pick grape bunches based on maturity is an important factor in ensuring their quality. However, the maturity of grape bunches in complex grape garden environments is easily affected by many factors. This paper proposes a grape maturity detection method based on You Only Look Once version (YOLOv) 8. The model's performance was improved by introducing a small object detection layer and replacing the faster implementation of the cross-stage partial bottleneck with 2 convolutions (C2f) module and the re-parameterized visual geometry group (RepVGG) module in the YOLOv8 network. A total of 1581 ‘Kyoho’ grape images were taken on-site with a mobile phone and divided into four maturity levels. Through ablation experiments, YOLOv8, adding a small object detection layer (YOLOv8-Head), replacing the C2f module with the RepVGG module (YOLOv8-Rep), and adding a small object detection layer and replacing the RepVGG module (YOLOv8-Rep-Head) were used for model training. The experimental results showed that the mean average precision 0.5 (mAP0.5) of the YOLOv8-Rep-Head model reaches 0.927, which is higher than 0.870 of YOLOv8, 0.880 of YOLOv8-Head, and 0.917 of YOLOv8-Rep. The precision (P) and recall (R) rates reached a high level, 0.902 and 0.861, respectively, and the balanced F (F1) score also increased to 0.88, registering the best performance of all the models. The experimental results show that the YOLOv8-Rep-Head model can accurately detect the maturity of ‘Kyoho’ grape bunches, provide reference significance for the intelligent picking of ‘Kyoho’ grape bunches, and have a positive impact on improving the commercial rate of ‘Kyoho’ grapes.