Growth characteristics based multi-class kiwifruit bud detection with overlap-partitioning algorithm for robotic thinning

文献类型： 外文期刊

第一作者： Dang, Haojie

作者： Dang, Haojie;He, Leilei;Shi, Yufei;Janneh, Lamin L.;Liu, Xiaojuan;Chen, Chi;Li, Rui;Kou, Xiaoxi;Fu, Longsheng;Ye, Hongbao;Chen, Jinyong;Fu, Longsheng;Fu, Longsheng;Fu, Longsheng;Janneh, Lamin L.;Majeed, Yaqoob

作者机构：

关键词： Main bud; Lateral bud; Deep learning; Overlap-partitioning algorithm; Different annotation strategies

期刊名称：COMPUTERS AND ELECTRONICS IN AGRICULTURE （ 影响因子：8.9； 五年影响因子：9.3 ）

ISSN： 0168-1699

年卷期： 2025 年 229 卷

页码：

收录情况： SCI

摘要： Bud thinning is a critical operation in the early stage of kiwifruit production, which is currently performed by skilled workers and has an urgent need to develop bud thinning robots. Accurate detection of kiwifruit buds is the first step, which focuses on distinguishing main bud and lateral bud. Kiwifruit buds are small, with similar shapes and colors in the main and lateral buds. Therefore, two kiwifruit bud detection methodologies were proposed to distinguish them. One is two-stage kiwifruit bud detection methodology (T-SKBDM) with an enhanced algorithm that leverage kiwifruit bud growth characteristics after network training for precise detection of main and lateral buds, and another is one-stage kiwifruit bud detection methodology (O-SKBDM) that classifies buds during the training. These methodologies adopted a two-classes annotation strategy (T-CAS) and a five-classes annotation strategy (F-CAS), respectively. In addition, both utilized an overlap-partitioning algorithm (OPA) that partitions large images into small images with overlapping areas. YOLOv8l model was trained on the dataset with different annotation strategies before and after using the OPA. Results showed that the T-CAS achieved a mean average precision (mAP) of 66.4 % before employing the OPA, which was 17.5 % higher than the F-CAS. With the OPA, mAPs of the T-CAS and F-CAS increased by 15.8 % and 17.7 %, respectively. Furthermore, T-SKBDM improved by 12.0 % and 14.3 % in distinguishing main and lateral buds, respectively, compared with the average precisions of 69.2 % and 66.2 % for O-SKBDM. These results indicate that the T-SKBDM assists in detecting kiwifruit buds and distinguishing the main and lateral buds, thus laying the foundation for robotic bud thinning.

分类号：

  • 相关文献
作者其他论文 更多>>

微信小程序