基于改进YOLOv5的小目标烟雾检测算法

doi:10.16265/j.cnki.issn1003-3033.2023.11.0854

摘要/Abstract

摘要：

为解决火灾中的小目标烟雾检测精度不高的问题,提出一种基于改进YOLOv5的小目标烟雾检测算法。首先,将特征融合注意力(FFA)模块引入至主干网络中,使模型专注于小目标烟雾特征信息的提取;其次,通过采用多尺度金字塔解耦头(MPDH)模块替换卷积层模块,以改进YOLOv5算法中预测头层的检测部分,用于提升小目标烟雾的定位精度;最后,在专有数据集上进行试验验证与分析。结果表明:基于改进的YOLOv5小目标烟雾检测算法在目标检测精度上达到85.4%,在准确率、召回率方面,相较于原始算法分别提高了3.2%、6.3%。

关键词: 小目标烟雾, 烟雾检测算法, YOLOv5, 特征融合注意力(FFA), 多尺度金字塔解耦头(MPDH)

Abstract:

In order to solve the problem of low accuracy of smoke detection for small target in fire, a smoke detection algorithm based on improved YOLOv5 was proposed. Firstly, FFA module was introduced into the backbone network, so that the model focused on the extraction of smoke feature information of small target. Secondly, the convolutional layer module was replaced by MPDH module to improve the detection part of the prediction head layer in the YOLOv5 algorithm, which was used to improve the positioning accuracy of small target smoke. Finally, the experimental results and analysis were carried out on the proprietary data set. The results show that the improved YOLOv5 small target smoke detection algorithm achieves 85.4% target detection accuracy, and the accuracy and recall rate are improved by 3.2% and 6.3% respectively compared with the original algorithm.

Key words: small target smoke, smoke detection algorithm, YOLOv5, feature fusion attention (FFA), multi-scale pyramid decoupling head (MPDH)

张军, 尹柳, 巩欣飞, 徐赫桦. 基于改进YOLOv5的小目标烟雾检测算法[J]. 中国安全科学学报, 2023, 33(11): 75-81.

ZHANG Jun, YIN Liu, GONG Xinfei, XU Hehua. Small target smoke detection algorithm based on improved YOLOv5[J]. China Safety Science Journal, 2023, 33(11): 75-81.

图/表 13

表1

图1

图2

图3

图4

图5

图6

图7

图8

表2

表3

图9

图10

参考文献 19

[1]	中国消防. 2021年消防接处警创新高,扑救火灾74.5万起[EB/OL]. [2022-01-20]. https://www.119.gov.cn/gk/sjtj/2022/26442.shtml.
[2]	国家消防救援局. 2022年全国警情与火灾情况[EB/OL]. [2023-03-24]. https://www.119.gov.cn/qmxfxw/xfyw/2023/36210.shtml.
[3]	张萌, 韩豫, 刘泽锋. 深度学习下建筑工人高空安全防护装备检测方法[J]. 中国安全科学学报, 2022, 32(5):140-146. doi: 10.16265/j.cnki.issn1003-3033.2022.05.1141
	ZHANG Meng, HAN Yu, LIU Zefeng. Detection method of high-altitude safety protective equipment for construction workers based on deep learning[J]. China Safety Science Journal, 2022, 32(5):140-146. doi: 10.16265/j.cnki.issn1003-3033.2022.05.1141
[4]	PUNDIR A S, RAMAN B. Deep belief network for smoke detection[J]. Fire Technology, 2017, 53(6):1943-1960. doi: 10.1007/s10694-017-0665-z
[5]	袁非牛, 夏雪, 李钢, 等. 面向烟雾识别与纹理分类的Gabor网络[J]. 中国图象图形学报, 2019, 24(2):269-281.
	YUAN Feiniu, XIA Xue, LI Gang, et al. GaborNet for smoke recognition and texture classification[J]. Journal of Image and Graphics, 2019, 24(2):269-281.
[6]	赵红成, 田秀霞, 杨泽森, 等. 改进YOLOv3的复杂施工环境下安全帽佩戴检测算法[J]. 中国安全科学学报, 2022, 32(5):194-200. doi: 10.16265/j.cnki.issn1003-3033.2022.05.0714
	ZHAO Hongcheng, TIAN Xiuxia, YANG Zesen, et al. Safety helmet wearing detection algorithm in complex construction environment based on improved YOLOv3[J]. China Safety Science Journal, 2022, 32(5):194-200. doi: 10.16265/j.cnki.issn1003-3033.2022.05.0714
[7]	张坚鑫, 郭四稳, 张国兰, 等. 基于多尺度特征融合的火灾检测模型[J]. 郑州大学学报:工学版, 2021, 42(5):13-18.
	ZHANG Jianxin, GUO Siwen, ZHANG Guolan, et al. Fire detection model based on multi-scale feature fusion[J]. Journal of Zhengzhou University: Engineering Science, 2021, 42(5):13-18.
[8]	REDMON J, DIVVALA S, GIRSHICK R, et al. You only look once: unified, real-time object detection[C]. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2016:779-788.
[9]	刘丽娟, 陈松楠. 一种基于改进SSD的烟雾实时检测模型[J]. 信阳师范学院学报:自然科学版, 2020, 33(2):305-311.
	LIU Lijuan, CHEN Songnan. Real-time smoke detection model based on improved SSD[J]. Journal of Xinyang Normal University: Natural Science Edition, 2020, 33(2):305-311.
[10]	谢书翰, 张文柱, 程鹏, 等. 嵌入通道注意力的YOLOv4火灾烟雾检测模型[J]. 液晶与显示, 2021, 36(10):1445-1453.
	XIE Shuhan, ZHANG Wenzhu, CHENG Peng, et al. Fire smoke detection model based on YOLOv4 with channel attention[J]. Chinese Journal of Liquid Crystals and Displays, 2021, 36(10):1445-1453. doi: 10.37188/CJLCD.2020-0312
[11]	汪常建, 丁勇, 卢盼成. 融合改进FPN与关联网络的Faster R-CNN目标检测[J]. 计算机工程, 2022, 48(2):173-179. doi: 10.19678/j.issn.1000-3428.0059979
	WANG Changjian, DING Yong, LU Pancheng. Object detection using Faster R-CNN combining improved FPN and relation network[J]. Computer Engineering, 2022, 48(2):173-179. doi: 10.19678/j.issn.1000-3428.0059979
[12]	LIU Wei, ANGUELOV D, ERHAN D, et al. SSD: single shot multibox detector[C]. Proceedings of the European Conference on Computer Vision, 2016:21-37.
[13]	REDMON J, FARHADI A. YOLO 9000: better, faster, stronger[C]. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017:7263-7271.
[14]	REDMON J, FARHADI A. YOLOv3: an incremental improvement[C]. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018:89-95.
[15]	付宏建, 白宏阳, 郭宏伟, 等. 融合多注意力机制的光学遥感图像目标检测方法[J]. 光子学报, 2022, 51(12):312-320.
	FU Hongjian, BAI Hongyang, GUO Hongwei, et al. Object detection method of optical remote sensing image with multi-attention mechanism[J]. Acta Photonica Sinica, 2022, 51(12):312-320.
[16]	张宸嘉, 朱磊, 俞璐. 卷积神经网络中的注意力机制综述[J]. 计算机工程与应用, 2021, 57(20):64-72. doi: 10.3778/j.issn.1002-8331.2105-0135
	ZHANG Chenjia, ZHU Lei, YU Lu. Review of attention mechanism in convolutional neural networks[J]. Computer Engineering and Applications, 2021, 57(20):64-72. doi: 10.3778/j.issn.1002-8331.2105-0135
[17]	孙泽强, 陈炳才, 崔晓博, 等. 融合频域注意力机制和解耦头的YOLOv5带钢表面缺陷检测[J]. 计算机应用, 2023, 43(1):242-249. doi: 10.11772/j.issn.1001-9081.2021111926
	SUN Zeqiang, CHEN Bingcai, CUI Xiaobo, et al. Strip steel surface defect detection by YOLOv5 algorithm fusing frequency domain attention mechanism and decoupled head[J]. Journal of Computer Applications, 2023, 43(1):242-249. doi: 10.11772/j.issn.1001-9081.2021111926
[18]	董刚, 谢维成, 黄小龙, 等. 深度学习小目标检测算法综述[J]. 计算机工程与应用, 2023, 59(11):16-27. doi: 10.3778/j.issn.1002-8331.2211-0377
	DONG Gang, XIE Weicheng, HUANG Xiaolong, et al. Review of small object detection algorithms based on deep learning[J]. Computer Engineering and Applications, 2023, 59(11):16-27. doi: 10.3778/j.issn.1002-8331.2211-0377
[19]	郑远攀, 许博阳, 王振宇. 改进的YOLOv5烟雾检测模型[J]. 计算机工程与应用, 2023, 59(7):214-221. doi: 10.3778/j.issn.1002-8331.2209-0421
	ZHENG Yuanpan, XU Boyang, WANG Zhenyu. Improved YOLOv5 smoke detection model[J]. Computer Engineering and Applications, 2023, 59(7):214-221. doi: 10.3778/j.issn.1002-8331.2209-0421

算法名称		核心思想	优势	不足
Faster R-CNN		引入区域生成网络生成候选目标区域,并通过共享特征提取和多任务损失函数实现目标的分类和位置回归	提出区域生成网络,检测精度提升	检测速度慢,小目标检测效果不佳
SSD		将目标检测任务转化为一个回归问题和一个分类问题,采用不同尺度和长宽比在不同尺度的特征图中进行目标检测	采用多尺度特征图进行检测,检测速度快	训练难度高,定位精度有限,小目标检测率低
算法名称		核心思想	优势	不足
YOLO	v1	将目标检测任务转化为一个回归问题,并通过单个神经网络同时进行目标的定位和分类	端到端训练,复杂性降低,检测速度快	密集目标检测率低,小目标检测效果差
	v2		引入多尺度训练,泛化能力强,检测效果好	使用预训练,难以迁移,训练速度慢
	v3		首次采用跨阶段局部网络结构,检测准确性提升	结构复杂度高,召回率差,小目标检测不精准
	v4		引入路径聚合网络,检测精度提高	复杂场景下误检率高,小目标的特征信息不易提取
	v5		不同场景适应能力强,检测速度快	小目标检测精度不准确,相互靠近的目标检测效果差

算法名称	mAP@.5/%	P/%	R/%	FPS
Faster R-CNN	85.2	86.4	79.5	46
SSD	78.5	77.2	72	238
YOLOv5	81.1	83.7	74.1	324
MAPDH	85.4	85.8	80.4	289

改进名称	FFA 模块	MPDH 模块	P/%	R/%	mAP@.5: .95/%
YOLOv5	×	×	83.7	74.1	38.6
改进1	√	×	82.9	79.4	40.6
改进2	×	√	84.5	77.2	40.8
MAPDH	√	√	85.8	80.4	43.2