基于改进FireNet的轻型火灾实时检测方法
Real-time fire detection method based on deep learning
刘庆 1聂晶 1王友军 1张坤 1李义1
作者信息
- 1. 贵州电网有限责任公司毕节供电局,贵州毕节 551700
- 折叠
摘要
针对火灾检测精确不高,时间长等问题.设计了基于改进FireNet的轻型火灾实时检测方法,通过获取视频图像数据,网络模型进行火灾分析和识别;首先,在FireNet特征提取阶段使用多尺度卷积网络并引入通道注意力机制,以提高回归精度.其次,对全连接层的神经元个数进行压缩优化,减少计算耗时.实验表明,改进的FireNet算法模型检测精度达到96.43％,模型存储空间0.96MB,检测帧率37.相比标准算法精度提高2.5％,存储空间压缩85％,检测帧率提升35％.
Abstract
In response to the issues of low accuracy and long processing time in fire detection,a lightweight real-time fire detection method based on the improved FireNet is designed.The method utilizes video image data to perform fire inference analysis and identification using a network model.Firstly,a multi-scale convolutional network is employed in the FireNet feature extraction stage and a channel attention mechanism is introduced to improve the regression accu-racy.Secondly,the number of neurons in the fully connected layer is compressed and optimized to reduce computational time.Experimental results show that the improved FireNet algorithm achieves a detection accuracy of 96.43％,with a model storage space of 0.96MB and a detection frame rate of 40.Compared to the standard algorithm,the improved method exhibits a 2.5％increase in accuracy,an 85％reduction in storage space,and a 40％improvement in detection frame rate.
关键词
火灾检测/卷积神经网络/多尺度卷积网络/注意力机制Key words
fire detection/convolutional neural network/multi-scale convolutional network/at-tention mechanism引用本文复制引用
基金项目
2022年贵州电网有限责任公司毕节供电局电力技术开发项目(0607002022030101SC00049)
出版年
2024
NETL
NSTL
万方数据