Computer Engineering and Applications ›› 2023, Vol. 59 ›› Issue (17): 286-294.DOI: 10.3778/j.issn.1002-8331.2302-0039

• Engineering and Applications • Previous Articles     Next Articles

Improved Coal Mine Smoke and Fire Detection Algorithm of YOLOv5s

LIU Chunxia, LI Chao, PAN Lihu, FAN Senlin   

  1. College of Computer Science and Technology, Taiyuan University of Science and Technology, Taiyuan 030024, China
  • Online:2023-09-01 Published:2023-09-01

改进YOLOv5s的煤矿烟火检测算法

刘春霞,李超,潘理虎,樊森霖   

  1. 太原科技大学 计算机科学与技术学院,太原 030024

Abstract: Aimed at the characteristics of long detection time and high false detection rate in traditional coal mine smoke and fire detection, an improved coal mine smoke and fire detection algorithm of YOLOv5s is proposed. First, in order to make up for the amount of parameters and calculations caused by subsequent improvements, GhostNet is used to reconstruct the neck of YOLOv5s to make the network more lightweight; second, in order to solve the problem of small target detection, the global context feature extraction module BoT3(bottleneck transformer) is proposed, which can provide global context information for small targets and help the model to predict them better. At the same time, an SA(shuffle attention) attention mechanism is added to each head position to make the model focus more on small target information and suppress noise interference; finally, the bounding box regression loss function CIoU is replaced with SIoU to improve regression accuracy and speed up the model of convergence. The experimental results show that the improved YOLOv5s has less parameters, less calculation, and higher accuracy, which can meet the requirements of coal mine pyrotechnic detection.

Key words: coal mine smoke and fire detection, YOLOv5s, shuffle attention(SA)attention mechanism, small target detection, SIoU

摘要: 针对传统的煤矿烟火检测存在检测时间长、误检率高等特点,提出了一种基于改进YOLOv5s的煤矿烟火检测算法。为了弥补后续改进带来的参数量和计算量等问题,使用GhostNet重构YOLOv5s的颈部,使得网络更加轻量化;为了解决小目标检测问题,提出了全局上下文特征提取模块BoT3(bottleneck transformer),该模块可以为小目标提供全局上下文信息,帮助模型更好地预测小目标。同时,在每个检测头位置加入SA(shuffle attention)注意力机制,使模型更聚焦小目标信息,抑制噪声的干扰;将边界框回归损失函数CIoU替换为SIoU,提高回归精度,加速模型的收敛。实验结果表明,改进之后的YOLOv5s参数量、计算量更小,精度更高,能够满足煤矿烟火检测要求。

关键词: 煤矿烟火检测, YOLOv5s, SA注意力机制, 小目标检测, SIoU