纪 超,侯 威,高鸣江,张 凡,杨 鹏,李小兵.基于 MBDC 和双重注意力的变电站人员穿戴检测[J].电子测量与仪器学报,2023,37(6):247-255 |
基于 MBDC 和双重注意力的变电站人员穿戴检测 |
Wear detection of substation staff based on MBDC and dual attention |
|
DOI: |
中文关键词: 变电站人员穿戴 多分支深度卷积 双重注意力机制 多通道交互注意力 高效通道注意力 |
英文关键词:wearing of substation personnel multi branch depth convolution dual attention mechanism multimodal interaction
attention efficient channel attention |
基金项目:陕西省重点研发计划(2020ZDLGY09-10)、金属挤压与锻造装备技术国家重点实验室开放课题(S2208100. W03)项目资助 |
|
|
摘要点击次数: 681 |
全文下载次数: 751 |
中文摘要: |
安全帽与工作服是变电站工作人员安全的重要保障,为解决现有检测模型对其检测精度低的问题,本文提出了 MBDC
和双重注意力的变电站人员穿戴检测算法。 该算法提出了多分支深度卷积(multi branch deep convolution, MBDC)网络增加深
度可分离卷积层以增强特征提取的完备性;然后提出多通道交互注意力(multimodal interaction attention, MIA)增加模型对小目
标的检测能力,并将 MIA 机制结合高效通道注意力( efficient channel attention, ECA)机制构成双重注意力机制,增强模型对于
小目标和遮挡目标的识别精度;最后引入焦点损失函数和 SIOU( scylla intersection over union)作为损失函数以解决正负样本不
平衡问题并加快收敛速度。 实验表明,本文算法全类平均精度达到 84. 88%,比原算法高 9. 92%,总体性能优于对比算法。 |
英文摘要: |
Helmets and work clothes are important guarantees for the safety of substation staff. In order to solve the problem of low
detection accuracy of existing detection models, this paper proposes a substation staff wear detection algorithm based on multi-branch
deep convolution and dual attention. The algorithm proposes a multi-branch deep convolution (MBDC) network to add a deep separable
convolution layer to enhance the completeness of feature extraction. Then, it is proposed that multimodal interaction attention (MIA) can
increase the detection ability of the model to small targets, and combine the MIA mechanism with efficient channel attention (ECA)
mechanism to form a dual attention mechanism to enhance the recognition accuracy of the model for small targets and occluded targets.
Finally, the focus loss function and SIOU ( scylla intersection over union) are introduced as the loss function to solve the problem of
positive and negative sample imbalance and accelerate the convergence speed. The experiment shows that the average accuracy of the
algorithm in this paper is 84. 88%, 9. 92% higher than the original algorithm, and the overall performance is better than the comparison
algorithm. |
查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|