基于全卷积回归分类的甲烷气体泄漏检测方法  

Methane Gas Leakage Detection Method Based on Total Convolutional Regression Classification

作　　者：黄任翔 蔺素珍[1] 李大威 和葆华 HUANG Ren-xiang;LIN Su-zhen;LI Da-wei;HE Bao-hua(College of Data Science and Technology,North University of China,Taiyuan Shanxi 030051,China;College of Electrical and Control Engineering,North University of China,Taiyuan Shanxi 030051,China;Shanxi Zilai Measurement and Control Technology CO.LTD,Taiyuan Shanxi 030000,China)

机构地区：[1]中北大学大数据学院,山西太原030051 [2]中北大学电气与控制工程学院,山西太原030051 [3]山西紫来测控技术有限公司,山西太原030000

出　　处：《计算机仿真》2024年第12期536-543,共8页Computer Simulation

基　　金：山西省研究生创新项目(2022Y630);山西省研究生创新项目(2022Y631);中北大学第十八届研究生科技立项项目(20221845);山西省专利转化专项计划项目(202302001)。

摘　　要：针对基于数学建模的红外气体检测方法无法解决气体逸散过程中虚景干扰的问题,提出一种基于全卷积回归和分类的甲烷气体检测方法。首先输入红外源图像,经过定向梯度特征图(Histogram of Oriented Gradient,HOG)提取模块,同时经过并行的ResNet18特征提取模块;其次将ResNet18特征与HOG特征融合;融合特征经过输出网络中的分类、回归模块,在分类模块中加入中心度偏差回归减少远离目标中心的稀薄气体、虚警产生的低质量边界框,进而提高网络模型的稳定性;最终分类模块提供的分类特征图、中心度偏差特征以及回归模块提供的边界框回归特征图经过边界框回归模块计算出最终的预测边界框。实验结果显示,检测准确率和重叠率分别可达到64.38%和51.38%,相比其它气体检测方法分别提升了20%~40%、10%~25%。表明本文方法可更为准确检测甲烷气体泄漏位置。This paper proposes a methane gas detection method based on fully convolutional regression and classification,which overcomes the problem of virtual scene interference in infrared gas detection methods based on mathematical modeling.Firstly,the input infrared source image is processed by a directional gradient feature map(Histogram of Oriented Gradient,HOG)extraction module,as well as a parallel ResNet18 feature extraction module.Secondly,the ResNet18 features are fused with the HOG features,and the fused feature is passed through the classification and regression modules in the output network.The classification module incorporates center deviation regression to reduce sparse gases far from the target center and low-quality boundary boxes that produce false alarms,thus improving the stability of the network model.Finally,the classification feature map,center deviation feature,and boundary box regression feature provided by the regression module are used to calculate the final predicted boundary box.Experimental results show that the detection accuracy and overlap rate can reach 64.38%and 51.38%,respectively,which improves the performance of other gas detection methods by 20%-40%and 10%-25%.This indicates that the proposed method can more accurately detect methane gas leakage locations.

关 键 词：目标检测 红外图像处理 多特征融合 深度学习 全卷积回归 

分 类 号：TP751.1[自动化与计算机技术—检测技术与自动化装置]