基于YOLOv5s的轻量化森林火灾检测算法研究

doi:10.11996/JG.j.2095-302X.2023010026

摘要/Abstract

摘要：

针对过去无人机搭载嵌入式设备巡检森林火灾精确率低、适应性差和软硬件限制高等问题，提出一种基于YOLOv5s的轻量化森林火灾目标检测算法。通过将YOLOv5s的骨干网络替换为轻量化网络Shufflenetv2，并以通道重组的思想，让骨干网络对图片信息的提取效率变得更快，在保持网络精度的同时保证检测速度；接着在Backbone与Neck的连接处加入为轻量化网络设计的CA位置注意力模块，可将图片不同的位置信息聚合到通道中，使被检对象关注度得以提高；最后在预测部分使用CIOU损失函数，能够更好的优化矩形框的长宽比和更快加速模型收敛。算法部署在嵌入式系统Jetson Xavier NX上的结果显示，改进后的网络模型大小与对比实验方法相比，最多减少了98%，准确率(Precision)达到92.6%，精确率(AP)达到95.3%，帧率(FPS)提升到132帧每秒，能满足在白天、黑夜或视野良好等情况下对森林火灾的实时性预防与检测，并具有良好的准确率和鲁棒性。

关键词: 目标检测, YOLOv5s, 轻量化, 位置注意力模块, 森林火灾检测

Abstract:

A new algorithm for light-weight forest fire object detection was proposed based on YOLOv5s to address the low accuracy, poor flexibility, and high software and hardware limitations of the previous UAV-embedded equipment for forest fire inspection. The proposed algorithm replaced the backbone of YOLOv5s with the light-weight network Shufflenetv2, employed the idea of channel recombination to improve the speed of the backbone network in picture information extraction, and maintained both high accuracy and fast detection speed. Then, a coordinate attention (CA) positional attention module specially designed for light-weight network was added to the connection between Backbone and Neck, which could aggregate different position information of pictures into the channel, thus improving the attention of the detected object. Finally, the CIOU loss function was utilized in the prediction part to better optimize the ratio of length to width of the rectangular frame and accelerate the model convergence. The results of the algorithm deployed on Jetson Xavier NX show that compared with the Faster-RCNN, SSD, YOLOv4, and YOLOv5s experimental methods, the improved network model size was reduced by up to 98%, increasing the precision to 92.6%, accuracy rate to 95.3%, and FPS to 132 frames/s. It can effectively achieve the real-time prevention and detection of forest fire in daylight, darkness, or good visibility, exhibiting good accuracy and robustness.

Key words: object detection, YOLOv5s, light-weight, positional attention module, forest fire detection

中图分类号:

TP391

皮骏, 刘宇恒, 李久昊. 基于YOLOv5s的轻量化森林火灾检测算法研究[J]. 图学学报, 2023, 44(1): 26-32.

PI Jun, LIU Yu-heng, LI Jiu-hao. Research on lightweight forest fire detection algorithm based on YOLOv5s[J]. Journal of Graphics, 2023, 44(1): 26-32.

图/表 11

参考文献 20

[1]	张彬彬, 帕孜来·马合木提. 基于YOLOv3改进的火焰目标检测算法[J]. 激光与光电子学进展, 2021, 58(24): 289-296.
	ZHANG B B, PAZILAI M H M T. Improved flame target detection algorithm based on YOLOv3[J]. Laser & Optoelectronics Progress, 2021, 58(24): 289-296 (in Chinese).
[2]	缪伟志, 陆兆纳, 王俊龙, 等. 基于视觉的火灾检测研究[J]. 森林工程, 2022, 38(1): 86-92, 100.
	MIAO W Z, LU Z N, WANG J L, et al. Fire detection research based on vision[J]. Forest Engineering, 2022, 38(1): 86-92, 100 (in Chinese).
[3]	DONG X D, YAN S, DUAN C Q. A lightweight vehicles detection network model based on YOLOv5[J]. Engineering Applications of Artificial Intelligence, 2022, 113: 104914. DOI URL
[4]	CHEN Z C, YANG J, CHEN L F, et al. Garbage classification system based on improved ShuffleNet v2[J]. Resources, Conservation and Recycling, 2022, 178: 106090. DOI URL
[5]	KRIZHEVSKY A, SUTSKEVER I, HINTON G E. ImageNet classification with deep convolutional neural networks[J]. Communications of the ACM, 2017, 60(6): 84-90. DOI URL
[6]	刘凯, 魏艳秀, 许京港, 等. 基于计算机视觉的森林火灾识别算法设计[J]. 森林工程, 2018, 34(4): 89-95.
	LIU K, WEI Y X, XU J G, et al. Design of forest fire identification algorithm based on computer vision full text replacement[J]. Forest Engineering, 2018, 34(4): 89-95 (in Chinese).
[7]	LOWE D G. Distinctive image features from scale-invariant keypoints[J]. International Journal of Computer Vision, 2004, 60(2): 91-110. DOI URL
[8]	PERRONNIN F, SÁNCHEZ J, MENSINK T. Improving the fisher kernel for large-scale image classification[M]//Computer Vision - ECCV 2010. Cham: Springer International Publishin, 2010: 143-156.
[9]	SZEGEDY C, LIU W, JIA Y Q, et al. Going deeper with convolutions[C]//2015 IEEE Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2015: 1-9.
[10]	仲亭玉, 刘文萍, 刘鹏举. 基于分数阶微分视频融合的森林烟火检测算法[J]. 北京林业大学学报, 2017, 39(3): 24-31.
	ZHONG T Y, LIU W P, LIU P J. A forest fire smoke detection algorithm based on fractional-calculus video fusion[J]. Journal of Beijing Forestry University, 2017, 39(3): 24-31 (in Chinese).
[11]	LIN T Y, GOYAL P, GIRSHICK R, et al. Focal loss for dense object detection[C]//2017 IEEE International Conference on Computer Vision. New York: IEEE Press, 2017: 2999-3007.
[12]	AVAZOV K, MUKHIDDINOV M, MAKHMUDOV F, et al. Fire detection method in smart city environments using a deep-learning-based approach[J]. Electronics, 2021, 11(1): 73. DOI URL
[13]	LIU S T, ZHANG N N, YU G. Lightweight security wear detection method based on YOLOv5[EB/OL]. [2022-02-20]. https://dl.acm.org/doi/10.1155/2022/1319029.
[14]	KRIZHEVSKY A, SUTSKEVER I, HINTON G E. ImageNet classification with deep convolutional neural networks[J]. Communications of the ACM, 2017, 60(6): 84-90. DOI URL
[15]	KRIZHEVSKY A, SUTSKEVER I, HINTON G E. ImageNet classification with deep convolutional neural networks[J]. Communications of the ACM, 2017, 60(6): 84-90. DOI URL
[16]	张苗, 李璞, 杨漪, 等. 基于目标检测卷积神经网络的图像型火灾探测算法[J]. 消防科学与技术, 2022, 41(6): 807-811.
	ZHANG M, LI P, YANG Y, et al. Image fire detection algorithms based on object detection convolutional neural networks[J]. Fire Science and Technology, 2022, 41(6): 807-811 (in Chinese).
[17]	DALAL N, TRIGGS B. Histograms of oriented gradients for human detection[J]. Proceedings of Computer Vision and Pattern Recognition, 2005, 1: 886-893.
[18]	PERRONNIN F, SÁNCHEZ J, MENSINK T. Improving the fisher kernel for large-scale image classification[M]//Computer Vision - ECCV 2010. Cham: Springer International Publishin, 2010: 143-156.
[19]	HOU Q B, ZHOU D Q, FENG J S. Coordinate attention for efficient mobile network design[EB/OL]. [2022-01-12]. https://arxiv.org/abs/2103.02907.
[20]	HE K M, ZHANG X Y, REN S Q, et al. Spatial pyramid pooling in deep convolutional networks for visual recognition[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2015, 37(9): 1904-1916. DOI PMID

Model summary	Layers	Parameters	Gradients	GFLOPS
CSPDarknet	272	7 233 211	7 233 211	16.4
Shuffnetv2	224	3 608 302	3 608 302	8.1

Model summary	Layers	Parameters	Gradients	GFLOPS
CSPDarknet	272	7 233 211	7 233 211	16.4
Shuffnetv2	224	3 608 302	3 608 302	8.1

Number	Shuffle (Backbone)	CIOU_Loss	SPPF	CA	Precison	Recall	AP
1	√	-	-	-	0.890	0.924	0.935
2	√	√	-	-	0.908	0.889	0.936
3	√	-	√	-	0.907	0.880	0.929
4	√	-	-	√	0.920	0.911	0.931
5	√	√	√	-	0.927	0.912	0.945
6	-	√	√	-	0.912	0.921	0.911
7	-	√	-	√	0.932	0.919	0.942
8	-	√	√	√	0.936	0.932	0.950
9	-	-	√	√	0.922	0.928	0.949
Ours	√	√	√	√	0.941	0.929	0.960

Number	Shuffle (Backbone)	CIOU_Loss	SPPF	CA	Precison	Recall	AP
1	√	-	-	-	0.890	0.924	0.935
2	√	√	-	-	0.908	0.889	0.936
3	√	-	√	-	0.907	0.880	0.929
4	√	-	-	√	0.920	0.911	0.931
5	√	√	√	-	0.927	0.912	0.945
6	-	√	√	-	0.912	0.921	0.911
7	-	√	-	√	0.932	0.919	0.942
8	-	√	√	√	0.936	0.932	0.950
9	-	-	√	√	0.922	0.928	0.949
Ours	√	√	√	√	0.941	0.929	0.960

算法	Backbone	模型大小(MB)	Precision	Recall	AP	FPS	检测用时(ms)
YOLOv4	CSPDarknet53	241.2	0.859	0.758	0.852	35	11.2
YOLOv4-tiny	CSPDarknet53-Tiny	25.3	0.869	0.792	0.843	42	4.5
YOLOv5s	CSPDarknet	11.5	0.923	0.909	0.944	79	9.8
YOLOv5s-P6	CSPDarknet	15.4	0.891	0.881	0.932	62	10.1
YOLOv5s-P7	CSPDarknet	15.5	0.906	0.892	0.939	65	11.3
YOLOv5s-Transformer	Transformer	20.1	0.915	0.911	0.950	70	9.2
SSD	VGGnet	180.3	0.613	0.521	0.631	22	60.7
Faster-RCNN	Resnet	212.3	0.732	0.612	0.756	28	53.1
Ours	Shufflenetv2	5.1	0.926	0.920	0.953	132	3.6