English

中国农机化学报

中国农机化学报 ›› 2023, Vol. 44 ›› Issue (4): 195-201.DOI: 10.13733/j.jcam.issn.2095-5553.2023.04.027

• 农业智能化研究 • 上一篇    下一篇

融合强化注意力机制的农田杂草识别方法

惠巧娟1,马伟2,边超1   

  1. 1. 银川科技学院信息工程学院,银川市,750021; 

    2. 宁夏大学新华学院,银川市,750021
  • 出版日期:2023-04-15 发布日期:2023-04-25
  • 基金资助:
    宁夏高等学校科学研究项目(NGY2020114)

Agricultural weeds recognition method based on enhanced attention mechanism

Hui Qiaojuan1, Ma Wei2, Bian Chao1   

  • Online:2023-04-15 Published:2023-04-25

摘要: 针对农作物与杂草交叉生长,导致杂草等目标难以识别的问题,提出一种融合强化注意力机制的农田杂草识别方法。首先,利用主干网络进行特征提取,并在此基础上提出一种强化注意力模块,从水平和垂直等两个维度细粒度进行位置特征编码,通过计算原始主干网络提取的特征与位置编码之间的偏移量,强化目标物体的定位与识别;然后,在单层注意力机制的基础上,引入上下文关系链条,进一步强化模型的泛化性能,最后,结合迁移学习的训练方式缓解小样本数据集极易造成过拟合的问题。通过测试单一目标物体和交叉生长的多目标物体在晴天、雨天和阴天等多场景环境下的识别性能,结果表明,本文方法分别可以实现单一目标物体和交叉生长的多目标物体92.84%和90.01%的平均识别准确率。

关键词: 农田杂草, 注意力机制, 强化注意力, 位置特征编码, 迁移学习

Abstract: Aiming at the problem that it is difficult to recognize weeds and other targets due to the cross growth between crops and weeds, an agricultural weeds recognition method based on enhanced attention mechanism is proposed. Firstly, the backbone network is used to extract the deep feature of the input images, and on this basis, an enhanced attention module is proposed, which is used to encode the position feature from the horizontal and vertical dimensions, and then strengthening the position and recognition of the target object by calculating the offset between the feature extracted from the original backbone network and the position encode. Then, the singlelayer based on attention mechanism, the context chain is introduced to further strengthen the generalization performance of the model. Finally, the training method combined with the transfer learning can alleviate the problem of the over fitting caused by small sample data sets. By testing the recognition performance of the single target object and cross growing multitarget objects in multi scene environments such as sunny, rainy and cloudy days, the results show that the proposed method can achieve the average recognition accuracy of 92.84% and 90.01% for the single target object and cross growing multitarget objects respectively.

Key words: agricultural weeds, attention mechanism, strengthen attention, position feature encode, transfer learning

中图分类号: