基于改进YOLOv4算法的煤矿火灾视频智能识别方法研究  

Research on intelligent video recognition method for coal mine fires based on improved YOLOv4 algorithm

作　　者：王伟峰 李煜 田丰[2] 张宝宝 何地 李高爽 李卓洋 WANG Weifeng;LI Yu;TIAN Feng;ZHANG Baobao;HE Di;LI Gaoshuang;LI Zhuoyang(College of Safety Science and Engineering,Xi’an University of Science and Technology,Xi’an,Shaanxi 710000,China;College of Communication and Information Technology,Xi’an University of Science and Technology,Xi’an,Shaanxi 710000,China;College of Electrical and Control Engineering,Xi’an University of Science and Technology,Xi’an,Shaanxi 710000,China)

机构地区：[1]西安科技大学安全科学与工程学院,陕西省西安市710000 [2]西安科技大学通信与信息工程学院,陕西省西安市710000 [3]西安科技大学电气与控制工程学院,陕西省西安市710000

出　　处：《中国煤炭》2025年第2期88-95,共8页China Coal

基　　金：陕西省重点研发计划项目(2021SF-472,2022QCY-LL-70);陕西省秦创原“科学家+工程师”队伍建设项目(2023KXJ-052)。

摘　　要：随着矿井智能化建设,煤矿火灾风险隐患逐渐增加。针对现有火灾检测算法存在准确率低以及对小火焰识别差的问题,提出一种煤矿火灾视频智能识别方法。该方法以YOLOv4为识别模型,采用群组归一化算法对模型归一化算法进行改进,并利用改进算法降低模型训练时批量值大小引起的误差;为降低矿井环境对火焰识别造成的火焰边缘信息损失,采用随机池化算法与SPP金字塔算法融合、深度可分离卷积与CSP算法融合,实现对动态演化的火焰进行跨尺度特征提取并融合、避免训练过程中的过拟合现象;为降低光源分布不均对视频火焰识别的影响,在模型中引入动态注意力机制,根据火灾视频识别信息的刺激强弱自动调整感受野大小。将标注后的火灾视频图像数据集输入到F YOLOv4算法模型进行训练及测试。结果表明,改进后的F YOLOv4火灾识别模型的平均检测精度达到97.3%左右,较原始模型提升了7.85%,表明该方法可提高检测速度和精度,可有效提高煤矿火灾识别的准确率。With the advancement of intelligent mine construction,the risk of coal mine fires has gradually increased.In response to the low accuracy of existing fire detection algorithms and their poor recognition of small flames,an intelligent recognition method for mine fire videos is proposed.This method utilizes YOLOv4 as the recognition model,improves the normalization algorithm by employing group normalization,which reduces errors caused by batch size during model training.To minimize the loss of flame edge information caused by the mine environment on flame recognition,a combination of random pooling and spatial pyramid pooling(SPP)algorithms,as well as depthwise separable convolution and cross stage partial(CSP)networks,is utilized to extract and fuse cross-scale features of dynamically evolving flames,thereby avoiding overfitting during training.To reduce the impact of uneven light source distribution on video flame recognition,a dynamic attention mechanism is introduced into the model,which automatically adjusts the receptive field size based on the intensity of stimuli from fire video recognition information.The fire video image data set is annotated and fed into the F YOLOv4 algorithm model for training and testing.The results show that the average detection accuracy of the improved F YOLOv4 fire recognition model is about 97.3%,an increase of 7.85%over the original model,indicating that this method can enhance detection speed and accuracy,effectively improving the accuracy of mine fire recognition.

关 键 词：YOLOv4 CSP改进 SPP改进 群组归一化 动态注意力机制 

分 类 号：TD752[矿业工程—矿井通风与安全] TP391.41[自动化与计算机技术—计算机应用技术]