计算机工程与应用 ›› 2022, Vol. 58 ›› Issue (17): 256-262.DOI: 10.3778/j.issn.1002-8331.2204-0498

• 图形图像处理 • 上一篇    下一篇

面向火焰快速检测的轻量化深度网络研究

王斌,李靖,赵康,周温   

  1. 1.中北大学 大数据学院,太原 030051 
    2.山西新思备科技股份有限公司,山西 晋中 030600
  • 出版日期:2022-09-01 发布日期:2022-09-01

Research on Lightweight Depth Network for Rapid Flame Detection

WANG Bin, LI Jing, ZHAO Kang, ZHOU Wen   

  1. 1.School of Data Science and Technology, North University of China, Taiyuan 030051, China
    2.Shanxi Xinsibei Technology Co., Ltd., Jinzhong, Shanxi 030600, China
  • Online:2022-09-01 Published:2022-09-01

摘要: 在易燃易爆场合火灾控制中火焰极速检测意义重大,其对算法实时性、准确度、抗干扰性有较高要求。为此提出一种基于改进YOLOv4-tiny轻量化抗干扰火焰检测深度网络。引入类火目标图像与真实火焰图像并通过Mosaic数据增强方式建立鲁棒性火焰检测数据集;对YOLOv4-tiny骨干网络采用深度可分离卷积进行改进,使得原网络更加轻量化;在特征金字塔网络FPN(feature pyramid network)中融合多尺度特征提高网络对多层特征的学习表示能力,并引出多检测头以适应火焰爆发过程中不同尺度火焰的精准检测;在FPN中引入ECA(efficient channel attention)通道注意力机制进一步提高检测精度。实验结果表明,提出的YOLOv4-tiny-L4参数量仅为4.22?MB,准确率高达94.1%,执行时间仅为46?ms,满足火焰快速检测基本要求。

关键词: 火焰检测, 轻量化网络, 改进YOLOv4-tiny, 注意力机制, 多尺度特征

Abstract: Extremely fast detection to flame is of great significance for fire control in flammable and explosive occasions. It has high requirements for real-time, accuracy and anti-interference of the algorithm. To this end, the lightweight anti-interference flame detection depth network based on improved YOLOv4-tiny is proposed. Firstly, the robust flame detection data set is established by Mosaic data enhancement and the pseudo target(image similar to flame). Secondly, the YOLOv4-tiny backbone network is improved by deep separable convolution, which makes the original network lighter. Then, these multi-scale features are fused in the feature pyramid network(FPN) to improve the learning and representation ability of the network for multi-layer features, and multiple detection heads are introduced to adapt to the accurate detection of different scale flames in the process of flame explosion. Finally, ECA(efficient channel attention) attention mechanism is introduced into FPN to further improve the detection accuracy. Experimental results show that the parameter quantity of YOLOv4-tiny-L4 is only 4.22?MB; the accuracy is as high as 94.1%; the execution time is only 46?ms, which meets the basic requirements of rapid flame detection.

Key words: flame detection, lightweight network, improved YOLOv4-tiny, attention mechanism, multiscale feature