«上一篇/Previous Article|本期目录/Table of Contents|下一篇/Next Article»

《郑州大学学报(工学版)》[ISSN:1671-6833/CN:41-1339/T]

卷:

42

期数:

2021年05期

页码:

13-18

栏目:

出版日期:

2021-09-10

文章信息/Info

Title:

Fire Detection Model ba＜x＞sed On Multi-scale Feature Fusion-ncig2020

作者:

张坚鑫; 郭四稳; 张国兰; 谭琳

广州大学计算机科学与网络工程学院;

Author(s):

Zhang Jianxin; Guo Si Jing; Zhang Guolan; Tan Lin;

School of Computer Science and Network Engineering, Guangzhou University;

关键词:

深度学习; 火灾检测; 卷积神经网络; 多尺度特征; 特征金字塔结构

Keywords:

Deep learning;  fire detection;  convolutional neural network;  multi -scale features;  feature pyramid structure

DOI:

10.13705/j.issn.1671-6833.2021.05.016

文献标志码:

A

摘要:

火灾的发生对于人民的生命及财产安全都会造成较大的危害，随着科技的进步，火灾检测设备与方法层出不穷。为了提高火灾检测的准确率，降低火灾的漏报率，使用图像型火灾检测方法逐渐成为主流的检测方法，深度学习的发展更是为火灾检测提供了新的思路与方法。改进的两场景检测模型 Faster R-CNN能够在火灾检测中发挥巨大的作用。首先使用 Resnet101 作为特征提取网络，接着使用特征金字塔结构 FPN 提取了 Resnet101 的浅层特征和高层特征，将 Resnet101 的浅层特征图输入 Inception Module 结构提取多种尺寸的卷积特征，最终使用像素注意力机制和信道注意力机制对目标位置进行强化并弱化其余部分，使得检测目标更加精确。该网络避免了主干网络特征提取不充分的问题，融合了多种尺度的特征来区分火灾区域和非火灾区域，有效提高了火灾图像数据集的检测准确率。该模型的平均检测准确率 MAP 相对于Resnet101 特征提取网络有了9个百分点的提升，相对于 Resnet101 加入特征金字塔结构网络 FPN 提升了5个百分点。该模型的检测效率与 Resnet101 特征提取网络、Resnet101 加入特征金字塔结构网络 FPN 都接近，检测一张火灾图像所需时间为0.86秒。该模型能够检测出室内和室外火灾的目标区域，精度相对原始模型得到较大的提升，效率与原始模型接近。

Abstract:

The occurrence of fire will cause great harm to peoples lives and property safety. With the advancement of technology, fire detection equipment and methods are emerging in endlessly.In order to improve the accuracy of fire detection and reduce the false alarm rate of fire, the use of image-ba＜x＞sed fire detection methods has gradually become the mainstream detection method, and the development of deep learning has provided new ideas and methods for fire detection. The improved two-scene detection model Faster R-CNN can play an important role in fire detection.First, Resnet101 is used as a feature extraction network, then the feature pyramid structure FPN is used to extract the shallow and high-level features of Resnet101, and the shallow feature map of Resnet101 is input Inception Module structure extracts convolution features of multiple sizes, and finally uses the pixel attention mechanism and channel attention mechanism to strengthen the target position and weaken the rest, making the detection target more accurate. This network avoids the problem of insufficient feature extraction of trunk network, and integrates features of various scales to distinguish fire area and non-fire area, which effectively improves the detection accuracy of fire image data sets. Compared with the Resnet101 feature extraction network, the average detection accuracy of the model MAP has been improved by 9 percentage points, and compared with the Resnet101 with feature pyramid network FPN, it has improved by 5 percentage points. The detection efficiency of this model is similar to that of Resnet101 feature extraction network and Resnet101 with feature pyramid network FPN , and the time of detecting one fire image is between 0.8 and 0.9 second. This model can detect the target area of indoor and outdoor fire, and the accuracy is greatly improved compared with the original model, and the efficiency is close to the original model.

常用功能

更新日期/Last Update: 2021-10-11