CTWheatNet: Accurate detection model of wheat ears in field

文献类型： 外文期刊

第一作者： Guan, Yujie

作者： Guan, Yujie;Pan, Jiaqi;Jia, Weikuan;Fan, Qingqi;Yang, Liangliang;Yin, Xiang;Jia, Weikuan

作者机构：

关键词： Object detection; GWHD-2021; Deep learning; Wheat ear detection; Dense detection

期刊名称：COMPUTERS AND ELECTRONICS IN AGRICULTURE （ 影响因子：7.7； 五年影响因子：8.4 ）

ISSN： 0168-1699

年卷期： 2024 年 225 卷

页码：

收录情况： SCI

摘要： Wheat ear detection is an indispensable step in wheat field yield estimation and is crucial for achieving scientific field management. However, wheat ears in the field vary in shape, frequent overlap, and are susceptible to illumination variations, posing significant challenges for wheat ear detection tasks. To meet the requirements of field yield estimation and improve wheat ear detection accuracy, this paper proposes the CTWheatNet model, which integrates CNN and Transformer to overcome their respective limitations in wheat ear detection tasks and achieve precise detection of wheat ears in the field. Firstly, a CNN feature extraction module is proposed to obtain multi-scale wheat ear features, avoiding the loss of wheat ear edge information caused by patch cutting. The Feature Processing Module is designed with a dual-branch structure to interactively fuse multi-scale features. Secondly, the output of the CNN feature extraction module is inputted into the Transformer feature extraction module to capture global wheat ear features. The Multi-Branch Self-Attention module of the dual-branch attention mechanism simultaneously focuses on low/high-resolution features to obtain both local dependencies and global features. The Quad Attention is proposed to reduce the computational cost of the selfattention module and improve the model's applicability. Finally, the proposed model is validated on the public wheat ear dataset (GWHD-2021). The experimental results show that CTWheatNet achieves a detection accuracy of 52.7% AP and 93.0% AP50. Compared with other advanced models, CTWheatNet also performs well. This model can provide technical support and theoretical reference for intelligent counting in the process of wheat field yield estimation and detection of other cereal crops.

分类号：

  • 相关文献
作者其他论文 更多>>

微信小程序