# 基于多尺度特征融合的火灾检测模型

(广州大学 计算机科学与网络工程学院，广东 广州 510006)

## 2 多尺度特征融合网络

Figure 1 Multi-scale feature fusion network structure

### 2.2 Inception Module网络结构

Figure 2 Inception Module network structure

### 2.3 多维注意网络结构

Figure 3 Structure of multidimensional attention network

### 2.4 损失函数

loss=lossrpn+lossfastrcnn+lossattention

(1)

RPN阶段的分类和回归损失如式(2)所示，其中分类损失lossr_c是判断方框中是否是火灾，也就是区分前景、背景两类物体的损失，如式(3)所示。lossr_l可以对方框位置进行评估和微调，也就是用于比较真实分类的预测参数和真实平移缩放参数的差别，如式(4)所示。lossr_l需要对损失进行L1平滑正则化处理，如式(5)所示，参数σ=1。Faster R-CNN阶段的分类损失计算和RPN阶段一致，而回归损失的L1平滑正则化参数σ=3。

lossrpn=lossr_c+lossr_l

(2)

(3)

(4)

(5)

(6)

Figure 4 Attention image

## 3 实验结果与分析

### 3.1 火灾检测实验数据集

Figure 5 Example drawing of annotation

### 3.2 仿真实验

3.2.1 实验过程

Figure 6 Loss function curve

3.2.2 检测精度评估

(7)

(8)

AP=PRdr

(9)

(10)

(11)

Table 1 Accuracy evaluation index

Figure 7 P-R curve

3.2.3 检测速度评估

Table 2 Performance comparison of different methods
in fire detection dataset

3.2.4 室内外火灾检测

Figure 8 Fire image input and output of indoor
and outdoor

## 4 结论

[1] 杨文柱,刘晴,王思乐,等.基于深度卷积神经网络的羽绒图像识别[J].郑州大学学报(工学版),2018,39(2):11-17.

[2] 黄文锋,徐珊珊,孙燚,等.基于多分辨率卷积神经网络的火焰检测[J].郑州大学学报(工学版),2019,40(5):80-84.

[3] FRIZZI S,KAABI R,BOUCHOUICHA M,et al.Convolutional neural network for video fire and smoke detection[C]//IECON 2016 - 42nd Annual Conference of the IEEE Industrial Electronics Society.Piscataway:IEEE,2016:877-882.

[4] SHARMA J,GRANMO O C,GOODWIN M,et al.Deep convolutional neural networks for fire detection in images[J].Engineering applications of neural networks,2017,744:183-193.

[5] SIMONYAN K，ZISSERMAN A.Very deep convolutional networks for large-scale image recognition[C]//International Conference on Learning Representations. San Diego,USA:ICLR,2015:1-14.

[6] HE K M,ZHANG X Y,REN S Q,et al.Deep residual learning for image recognition[C]//2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR).Piscataway:IEEE,2016:770-778.

[7] 林作永,谌瑶.基于深度卷积神经网络的火灾预警算法研究[J].信息通信,2018(5):38-42.

[8] SZEGEDY C,VANHOUCKE V,IOFFE S,et al.Rethinking the inception architecture for computer vision[C]//2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR).Piscataway:IEEE,2016:2818-2826.

[9] HOWARD A G,ZHU M L,CHEN B,et al.MobileNets:efficient convolutional neural networks for mobile vision applications[EB/OL].(2017-04-17)[2020-12-01].https://arxiv.org/pdf/1704.04861.pdf.

[10] REN S Q,HE K M,GIRSHICK R,et al.Faster R-CNN:towards real-time object detection with region proposal networks[J].IEEE transactions on pattern analysis and machine intelligence,2017,39(6):1137-1149.

[11] LIU W,ANGUELOV D,ERHAN D,et al.SSD:single shot MultiBox detector[M]//Computer Vision-ECCV 2016.Cham:Springer International Publishing,2016:21-37.

[12] DAI J F,LI Y,HE K M,et al.R-FCN:object detection via region-based fully convolutional networks[C]//NIPS′16:Proceedings of the 30th International Conference on Neural Information Processing Systems.Barcelona,Spain:NIPS,2016:379-387.

[13] IOFFE S,SZEGEDY C.Batch normalization:accelerating deep network training by reducing internal covariate shift[C]//32nd International Conference on Machine Learning 2015.Lille,France:ICML,2015:448-456.

[14] 夏雪,袁非牛,章琳,等.从传统到深度:视觉烟雾识别、检测与分割[J].中国图象图形学报,2019,24(10):1627-1647.

[15] LIN T Y,DOLLR P,GIRSHICK R,et al.Feature pyramid networks for object detection[C]//2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR).Piscataway:IEEE,2017:936-944.

[16] SZEGEDY C,LIU W,JIA Y Q,et al.Going deeper with convolutions[C]//2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR).Piscataway:IEEE,2015:1-9.

[17] SZEGEDY C，IOFFE S，VANHOUCKE V，et al. Inception-v4，Inception-ResNet and the impact of residual connections on learning[EB/OL].(2016-08-23)[2020-12-01].https://arxiv.org/pdf/1602.07261.pdf.

[18] HU J,SHEN L,SUN G.Squeeze-and-excitation networks[C]//2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition.Piscataway:IEEE,2018:7132-7141.

[19] REDMON J，FARHADI A.YOLOv3:an incremental improvement[EB/OL].(2018-04-08)[2020-12-01]. https://arxiv.org/pdf/1804.02767.pdf.

# Fire Detection Model Based on Multi-scale Feature Fusion

ZHANG Jianxin, GUO Siwen, ZHANG Guolan, TAN Lin

(School of Computer Science and Cyber Engineering, Guangzhou University, Guangzhou 510006, China)

Abstract: This paper aims to modify the two-scene detection model Faster R-CNN. Specifically, this model uses Resnet101 to extract features which are processed by pyramid structure FPN to extract the shallow and high-level features of Resnet101. The shallow feature map of Resnet101 is input into Inception Module structure to obtain the convolutional features of multiple sizes, and finally the proposed model uses the pixel attention mechanism and channel attention mechanism to emphasize the target position and weaken the rest, which makes 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, thus effectively improves the detection accuracy of fire image data sets, and mean average precision MAP is 0.851.

Key words: deep learning; fire detection; convolutional neural network; multi-scale features; feature pyramid network

doi:10.13705/j.issn.1671-6833.2021.05.016