基于改进YOLOv5的火焰检测方法

doi:10.19678/j.issn.1000-3428.0065025

摘要/Abstract

摘要：

现有基于图像的火焰检测方法难以兼顾实时性和准确性，且缺乏对小火焰目标精准识别的能力，无法有效应对小火点灭火等应用场景。YOLOv5算法与传统主流算法相比在检测的实时性上有很大优势，为提升火焰检测精度，提出一种基于改进YOLOv5的火焰实时检测方法。针对YOLOv5模型进行改进：在特征提取部分嵌入协同注意力机制模块，在不损失特征信息的情况下减少特征冗余，以帮助模型更精确地定位火焰特征；在特征融合部分增加一个专门针对小火焰目标的检测层，并添加对应的特征提取及特征融合模块，以帮助模型有效获取感受野小于8×8像素的火焰特征；在损失函数的计算部分使用α-CIoU作为新的边界框损失函数，以提升模型的收敛速度和对小数据集的鲁棒性。此外，通过模型预训练和迁移学习的方法对火焰检测模型各层结构的权重参数进行初始化，防止梯度消失，提升训练效果。实验结果表明，改进后的火焰检测模型检测精度为96.6%，较YOLOv5原始模型提升7.4个百分点，并且检测速度达到68帧/s，模型大小仅15.4 MB，在大幅提升精度的基础上能够同时满足消防灭火机器人对火焰检测实时性和轻量化的要求。

关键词: 火焰检测, 注意力机制, 特征融合, YOLOv5算法, 边界损失函数

Abstract:

The existing image-based flame detection approach finds it challenging to balance real-time and precision, and it is incapable of accurately identifying small flame targets, making it ineffective for application situations such as small fire extinguishing. In terms of real-time detection, the YOLOv5 algorithm provides significant benefits over conventional techniques. A real-time flame detection method based on improved YOLOv5 is proposed to increase flame detection accuracy. First, to help the model locate the flame features more accurately, a coordinate attention mechanism module is embedded in the feature extraction portion of the YOLOv5 model.This module can reduce feature redundancy without sacrificing the feature information. Second, to help the model successfully obtain flame features with a receptive field smaller than 8×8 pixels, a detection layer specifically designed for small flame targets is added to the feature fusion portion of the algorithm along with the corresponding feature extraction and feature fusion modules. Finally, to increase the model's speed of convergence and robustness to small datasets, α-CIoU is employed as a new bounding box loss function in the computation phase of the loss function.Additionally, model pretraining and transfer learning techniques are used to initialize the weight parameters of each layer structure of the flame detection model to prevent the gradient from dissipating and enhance the training effect. According to the experimental findings, the proposed flame detection model shows an accuracy rate of 96.6%, which is 7.4 percentage points higher than that of the YOLOv5 original model.Additionally, the detection speed of this model is 68 frame/s, and its size is only 15.4 MB. On the basic of significantly improving accuracy, it can also meet the requirements of firefighting robots for real-time and lightweight flame detection.

Key words: flame detection, attention mechanism, feature fusion, YOLOv5 algorithm, boundary loss function

陈露萌, 曹彦彦, 黄民, 谢鑫钢. 基于改进YOLOv5的火焰检测方法[J]. 计算机工程, 2023, 49(8): 291-301, 309.

Lumeng CHEN, Yanyan CAO, Min HUANG, Xingang XIE. Flame Detection Method Based on Improved YOLOv5[J]. Computer Engineering, 2023, 49(8): 291-301, 309.

https://www.ecice06.com/CN/Y2023/V49/I8/291

图/表 15

图1 YOLOv5原始网络结构

Fig.1 Original network structure of YOLOv5

图2 YOLOv5子模块结构

Fig.2 Submodule structure of YOLOv5

图3 CA模块结构

Fig.3 Structure of CA module

图4 改进后的YOLOv5整体网络结构

Fig.4 Overall network structure of improved YOLOv5

图5 GIoU相同的3种情况

Fig.5 Three cases when GIoU is the same

图6 火焰检测方法的整体流程

Fig.6 Overall process of flame detection method

图7 LabelImg图片标注

Fig.7 LabelImg image annotation

图8 火焰图像的数据增强

Fig.8 Data augmentation of flame image

图9 Mosaic数据增强

Fig.9 Mosaic data augmentation

图10 数据集目标框分布情况

Fig.10 Dataset target box distribution

图11 mAP(0.5)对比

Fig.11 mAP(0.5) comparison

图12 损失函数对比

Fig.12 Comparison of loss functions

表1 原始网络与采用不同改进策略的网络的实验结果对比

Table 1 Comparison of experimental results between the original network and the networks with different improvement strategies

网络	改进策略				mAP(0.5)/%	检测速度/(帧·s^-1)
网络	协同注意力机制	小尺度检测层	α-CIoU损失函数	迁移学习	mAP(0.5)/%	检测速度/(帧·s^-1)
YOLOv5s原始网络	×	×	×	×	89.2	71
改进网络1	√	×	×	×	90.3	70
改进网络2	√	√	×	×	93.7	63
改进网络3	√	√	√	×	95.2	65
改进网络4	√	√	√	√	96.6	68

表2 与主流目标检测模型的性能对比

Table 2 Performance comparison with mainstream target detection models

模型	精确率/%	召回率/%	mAP(0.5)/%	mAP(0.5∶0.95)/%	检测速度/(帧·s^-1)	模型大小/MB
CenterNet	82.1	80.6	83.8	66.8	21	88.0
Faster R-CNN	83.7	84.3	87.6	70.7	6	108.0
YOLOv3	82.6	83.5	85.7	68.4	52	235.0
YOLOv4	90.7	92.3	92.7	82.1	54	244.0
YOLOv5s	89.1	86.9	89.2	72.5	71	14.0
Fire-YOLO	88.9	89.7	90.3	79.3	42	49.6
本文改进模型	93.8	94.3	96.6	85.7	68	15.4

图13 火焰检测的测试结果

Fig.13 Flame detection test results

参考文献 27

1	苏展, 张防震, 王瑞, 等. 基于RGB+HSI颜色模型的早期火焰检测算法研究. 河南教育学院学报(自然科学版), 2021, 30(3): 28- 32. URL
	SU Z, ZHANG F Z, WANG R, et al. Research on early flame detection algorithm based on RGB+HSI color model. Journal of Henan Institute of Education(Natural Science Edition), 2021, 30(3): 28- 32. URL
2	曲娜, 王建辉, 陈琦华. 基于YCbCr颜色空间和BP神经网络的火焰图像检测. 沈阳大学学报(自然科学版), 2019, 31(4): 298- 301. URL
	QU N, WANG J H, CHEN Q H. A flame image detection method based on YCbCr color space and BP neural network. Journal of Shenyang University(Natural Science), 2019, 31(4): 298- 301. URL
3	杨述斌, 周敏瑞, 潘伟. 基于烟雾图像动态多帧差分法的火灾识别研究. 自动化与仪表, 2021, 36(7): 47-50, 100 URL
	YANG S B, ZHOU M R, PAN W. Research on fire recognition based on smoke image dynamic multi-frame difference method. Automation & Instrumentation, 2021, 36(7): 47-50, 100 URL
4	苗续芝, 陈伟, 毕方明, 等. 基于改进FOA-SVM的矿井火灾图像识别. 计算机工程, 2019, 45(4): 267- 274. URL
	MIAO X Z, CHEN W, BI F M, et al. Mine fire image recognition based on improved FOA-SVM. Computer Engineering, 2019, 45(4): 267- 274. URL
5	何欣. 基于深度学习的火焰识别检测研究[D]. 北京: 北京邮电大学, 2021.
	HE X. Research on frame recognition and detection based on deep learning[D]. Beijing: Beijing University of Posts and Telecommunications, 2021. (in Chinese)
6	朱晓妤, 严云洋, 刘以安, 等. 基于深度森林模型的火焰检测. 计算机工程, 2018, 44(7): 264- 270. URL
	ZHU X Y, YAN Y Y, LIU Y A, et al. Flame detection based on deep forest model. Computer Engineering, 2018, 44(7): 264- 270. URL
7	SUN X Q, XIN S J. Flame detection algorithm based on YOLOv4[C]//Proceedings of the 7th International Conference on Intelligent Computing and Signal Processing. Washington D. C., USA: IEEE Press, 2022: 314-317.
8	赵媛媛, 朱军, 谢亚坤, 等. 改进Yolo-v3的视频图像火焰实时检测算法. 武汉大学学报(信息科学版), 2021, 46(3): 326- 334. URL
	ZHAO Y Y, ZHU J, XIE Y K, et al. A real-time video flame detection algorithm based on improved Yolo-v3. Geomatics and Information Science of Wuhan University, 2021, 46(3): 326- 334. URL
9	王国平, 严云洋, 高尚兵, 等. 基于Fire-YOLO的火焰检测算法. 信息与电脑, 2022, 34(5): 49- 52. URL
	WANG G P, YAN Y Y, GAO S B, et al. Flame detection algorithm based on Fire-YOLO. China Computer & Communication, 2022, 34(5): 49- 52. URL
10	胡燕, 王慧琴, 黄东宇, 等. 基于改进GA-RS的火焰图像特征自适应选择. 计算机工程, 2015, 41(8): 186- 189. URL
	HU Y, WANG H Q, HUANG D Y, et al. Adaptive selection of flame image features based on improved GA-RS. Computer Engineering, 2015, 41(8): 186- 189. URL
11	REN S Q, HE K M, GIRSHICK R, et al. Faster R-CNN: towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017, 39(6): 1137- 1149.
12	HE K M, GKIOXARI G, DOLLÁR P, et al. Mask R-CNN[C]//Proceedings of IEEE International Conference on Computer Vision. Washington D. C., USA: IEEE Press, 2017: 2980-2988.
13	CAI Z W, VASCONCELOS N. Cascade R-CNN: delving into high quality object detection[C]//Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition. Washington D. C., USA: IEEE Press, 2018: 6154-6162.
14	LIN T Y, GOYAL P, GIRSHICK R, et al. Focal loss for dense object detection[C]//Proceedings of IEEE International Conference on Computer Vision. Washington D. C., USA: IEEE Press, 2017: 2999-3007.
15	DUAN K W, BAI S, XIE L X, et al. CenterNet: keypoint triplets for object detection[C]//Proceedings of IEEE/CVF International Conference on Computer Vision. Washington D. C., USA: IEEE Press, 2020: 6568-6577.
16	REDMON J, DIVVALA S, GIRSHICK R, et al. You only look once: unified, real-time object detection[C]//Proceedings of IEEE Conference on Computer Vision and Pattern Recognition. Washington D. C., USA: IEEE Press, 2016: 779-788.
17	REDMON J, FARHADI A. YOLO9000: better, faster, stronger[C]//Proceedings of IEEE Conference on Computer Vision and Pattern Recognition. Washington D. C., USA: IEEE Press, 2017: 6517-6525.
18	REDMON J, FARHADI A. YOLOv3: an incremental improvement[EB/OL]. [2021-10-26]. https://arxiv.org/abs/1804.02767.
19	LIN T Y, DOLLÁR P, GIRSHICK R, et al. Feature pyramid networks for object detection[C]//Proceedings of IEEE Conference on Computer Vision and Pattern Recognition. Washington D. C., USA: IEEE Press, 2017: 936-944.
20	BOCHKOVSKIY A, WANG C Y, LIAO H Y M. YOLOv4: optimal speed and accuracy of object detection[EB/OL]. [2021-10-26]. https://arxiv.org/abs/2004.10934.
21	WANG C Y, MARK LIAO H Y, WU Y H, et al. CSPNet: a new backbone that can enhance learning capability of CNN[C]//Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition. Washington D. C., USA: IEEE Press, 2020: 1571-1580.
22	HE K M, ZHANG X Y, REN S Q, et al. Spatial pyramid pooling in deep convolutional networks for visual recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2015, 37(9): 1904- 1916.
23	LIU S, QI L, QIN H F, et al. Path aggregation network for instance segmentation[C]//Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition. Washington D. C., USA: IEEE Press, 2018: 8759-8768.
24	REZATOFIGHI H, TSOI N, GWAK J, et al. Generalized intersection over union: a metric and a loss for bounding box regression[C]//Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition. Washington D. C., USA: IEEE Press, 2020: 658-666.
25	HOU Q B, ZHOU D Q, FENG J S. Coordinate attention for efficient mobile network design[C]//Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition. Washington D. C., USA: IEEE Press, 2021: 13708-13717.
26	ZHENG Z H, WANG P, REN D W, et al. Enhancing geometric factors in model learning and inference for object detection and instance segmentation[EB/OL]. [2021-10-26]. https://arxiv.org/abs/2005.03572.
27	HE J B, ERFANI S, MA X J, et al. Alpha-IoU: a family of power intersection over union losses for bounding box regression[EB/OL]. [2021-10-26]. https://arxiv.org/abs/2110.13675.

[1]	王炼红, 林飞鹏, 李潇瑶, 谌桂枝, 周莉. 融入课程知识图谱的KMAKT预测[J]. 计算机工程, 2024, 50(7): 23-31.
[2]	谭巨全, 王然. 特征融合下田径录像3D人体动作DTW捕捉算法[J]. 计算机工程, 2024, 50(7): 71-78.
[3]	钱来, 赵卫伟. 基于对比学习和注意力机制的文本分类方法[J]. 计算机工程, 2024, 50(7): 104-111.
[4]	刘建敏, 林晖, 汪晓丁. 基于图注意力机制的无地图场景轨迹预测方法[J]. 计算机工程, 2024, 50(7): 144-153.
[5]	屠乃威, 焦猛, 阎馨. 复杂环境下输电线路鸟巢目标图像检测模型[J]. 计算机工程, 2024, 50(7): 216-226.
[6]	张溢文, 蔡满春, 陈咏豪, 朱懿, 姚利峰. 融合空间特征的多尺度深度伪造检测方法[J]. 计算机工程, 2024, 50(7): 240-250.
[7]	肖慈, 徐杨, 张永丹, 冯明文, 黄易仟. 结合注意力和低光增强的夜间语义分割[J]. 计算机工程, 2024, 50(7): 271-281.
[8]	张锡英, 孙守东, 于海浩, 边继龙. 基于空间传播的多视图三维重建[J]. 计算机工程, 2024, 50(7): 293-302.
[9]	王晋涛, 秦昂, 张元, 陈一飞, 王廷凤, 谢承霖, 邹刚. 基于注意力增强与特征融合的中文医学实体识别[J]. 计算机工程, 2024, 50(7): 324-332.
[10]	贵向泉, 刘世清, 李立, 秦庆松, 李唐艳. 基于改进YOLOv8的景区行人检测算法[J]. 计算机工程, 2024, 50(7): 342-351.
[11]	林芷薇, 杨祖元, 王斯秋, 杨超. 基于多尺度线性全局注意力的运动员检测算法[J]. 计算机工程, 2024, 50(7): 352-359.
[12]	李亚康, 陈刚. 小角中子散射物理模型自动化筛选[J]. 计算机工程, 2024, 50(6): 56-64.
[13]	程腾腾, 姚春龙, 于晓强, 李旭, 王庆丰. 基于多头注意力机制融合常识知识的共情对话生成[J]. 计算机工程, 2024, 50(6): 94-101.
[14]	更藏措毛, 黄鹤鸣, 杨毅杰. 融合多尺度特征与上下文信息的语音增强方法[J]. 计算机工程, 2024, 50(6): 138-147.
[15]	贺姗, 蔺素珍, 王彦博, 李大威. 基于特征融合的多波段图像描述生成方法[J]. 计算机工程, 2024, 50(6): 236-244.

选择文件类型/文献管理软件名称

选择包含的内容