摘要
为了满足火灾实时蔓延对检测速度和准确率的更高要求,在YOLOv4的基础上,提出了一种轻量化火灾检测方法。将改进的MobileNetV3作为主干特征提取网络来降低模型复杂度,提高火灾检测速度,并引入高效通道注意力(ECA)机制模块,有效地捕获了跨通道交互,增强了对火灾目标区域的重点关注。在加强特征提取网络部分,采用了加权双向特征金字塔(BiFPN)网络结构,不增加额外参数的同时,融合了更多不同尺度的特征,对不同大小火灾区域的检测精度有了显著提升。实验结果表明:所提方法具有较好的火灾检测效果,在自建的数据集上,平均精度达到了86.4%,检测速度达到了58 fps,相较于原模型,分别提升了2.3%和24 fps,同时模型大小缩减了79%。
In order to meet the higher requirements of detection speed and accuracy of real-time spread of fire,a lightweight fire detection method based on YOLOv4 is proposed.The improved MobileNetV3 is used as the backbone feature extraction network to reduce the complexity of the model and increase the speed of fire detection,and introduce the efficient channel attention(ECA)mechanism module,which effectively captures cross-channel interaction and enhances the focus on the fire target area.In the enhanced feature extraction network part,the BiFPN weighted two-way feature pyramid structure is adopted,which fuses more features of different scales without adding additional parameters,and the detection precision of fire areas of different sizes has been significantly improved.The experimental results show that the proposed method has a good fire detection effect.On the self-built dataset,the average precision reaches 86.4%,and the detection speed reaches 58 fps.Compared with the original model,it is improved by 2.3%and 24 fps,and at the same time,the size of the model is reduced by 79%.
作者
郝新泽
施一萍
邓源
秦瑶
刘瑾
HAO Xinze;SHI Yiping;DENG Yuan;QIN Yao;LIU Jin(School of Electronic and Electrical Engineering,Shanghai University of Engineering and Technology,Shanghai 201620,China)
出处
《传感器与微系统》
CSCD
北大核心
2023年第8期143-147,共5页
Transducer and Microsystem Technologies
基金
国家自然科学基金资助项目(61701296)。
关键词
火灾检测
深度学习
注意力机制
多尺度特征融合
深度可分离卷积
fire detection
deep learning
attention mechanism
multi-scale feature fusion
depthwise separable convolution
作者简介
郝新泽(1998-),男,硕士研究生,研究方向为深度学习和目标检测;通讯作者:施一萍(1964-),女,副教授,研究领域为深度学习和智能控制等。
