改进Unet网络的汽油管道泄漏高光谱图像检测  

Hyperspectral Image Detection of Gasoline Pipeline Leakage Using Improved Unet Network

作　　者：王克明 公维佳 王海明 蔡永军 刘嘉星 孙磊 宋丽梅[1] 李金义[1] WANG Ke-ming;GONG Wei-jia;WANG Hai-ming;CAI Yong-jun;LIU Jia-xing;SUN Lei;SONG Li-mei;LI Jin-yi(Tianjin Key Laboratory of Intelligent Control of Electrical Equipment,School of Control Science and Engineering,Tiangong University,Tianjin 300387,China;Science and Technology Research Institute Branch,National Oil and Gas Pipeline Network Group Co.,Ltd.,Tianjin 300450,China;Northern Pipeline LLC,National Oil and Gas Pipeline Network Group Co.,Ltd.,Langfang 065000,China;Shandong Branch,National Oil and Gas Pipeline Network Group Co.,Ltd.,Jinan 250000,China)

机构地区：[1]天津工业大学控制科学与工程学院天津市电气装备智能控制重点实验室,天津300387 [2]国家石油天然气管网集团有限公司科学技术研究总院分公司,天津300450 [3]国家石油天然气管网集团有限公司北方管道有限责任公司,河北廊坊065000 [4]国家石油天然气管网集团有限公司山东省分公司,山东济南250000

出　　处：《光谱学与光谱分析》2025年第5期1476-1484,共9页Spectroscopy and Spectral Analysis

基　　金：国家自然科学基金项目(62375204);天津市科技计划项目(23YDTPJC00170);天津市自然科学基金项目(20JCYBJC00160)资助。

摘　　要：针对当前汽油管道泄漏检测效率低、无法精准分割泄露区域边缘的局限性,提出一种基于高光谱图像与深度学习结合的汽油管道泄漏检测方法。首先进行两种型号汽油在土壤、水背景下的特征光谱波段提取;利用连续投影算法实现汽油高光谱图像数据降维;将汽油反射率作为输入,均方根误差为回归参数获得汽油反射峰附近的18个特征波段;采用图像旋转角度、横向或纵向翻转、在图像中注入随机噪声等方式实现数据集样本扩充。其次对Unet高光谱图像语义分割模型进行改进,将Unet网络编码器部分替换成密集连接模块加强各层级间的信息交流,减轻计算量提高模型检测速度;引入通道注意力机制模块,使模型对汽油图像空间和光谱层面两特征信息同时关注,提高模型检测精度;引入失活层的概念,通过暂时关闭网络中的一部分神经元降低网络的复杂性,同时在训练过程中设置适当的时间点实施早停策略从而防止过拟合。最后进行了消融实验和对比实验。消融实验结果验证了密集连接模块和通道注意力机制模块对提高网络分割精度和召回率的有效性;在自建数据集上的定量对比实验结果表明,模型对滴落汽油的分割精度为90.34%,平均每张图片检测时间为0.23 s,与Unet、PSE-Unet和HLCA-Unet模型相比,平均准确率分别增加了14.39%、8.01%和2.73%,召回率分别增加了8.95%、8.02%和6.55%,测试时间与Unet、PSE-Unet模型相比分别减少了10.83%和16.97%,检测优越性定性体现在泄露油滴与背景交会的轮廓更符合原图,本模型可以获得更加准确的汽油特征信息,为汽油管道泄漏检测提供了新的技术方案。此外,在公开的Pavia University遥感数据集上与当前Unet、PSE-Unet、HLCA-Unet模型检测进行对比,模型仍表现出更好的分割效果,体现出较强的普适性和泛化能力,可用于多种类型的高光谱图像语义分割。In view of the limitations of the low efficiency of gasoline pipeline leak detection and the inability to accurately segment the edge of the leak region,a gasoline pipeline leak detection method is proposed based on hyperspectral image and deep learning.Firstly,the characteristic spectral bands of the two types of gasoline under the background of soil and water were extracted.The continuous projection algorithm was used to reduce the dimensionality of gasoline hyperspectral image data.The gasoline reflectivity was taken as input,and the root-mean-square error was the regression parameter used to obtain 18 characteristic bands near the gasoline reflection peak.Image rotation Angle,horizontal or vertical inversion,and random noise injection into the image are used to expand the dataset sample.Secondly,the Unet hyperspectral image semantic segmentation model is improved,and the network encoder part of Unet is replaced with a dense connection module to strengthen the information exchange between different levels,reduce the computational load,and improve the model detection speed.The spectral attention mechanism module is introduced to make the model pay attention to gasoline image space and spectral features and improve the model detection accuracy.The concept of an inactivation layer is introduced to reduce the complexity of the network by temporarily shutting down some neurons in the network.At the same time,an appropriate time point is set in the training process to implement the early stop strategy to prevent overfitting.Finally,the ablation experiment and comparison experiment were carried out.The results of ablation experiments validate the effectiveness of the dense connection module and the spectral attention mechanism module in improving the network's segmentation accuracy and recall rate.Quantitative comparison experiments on self-built data sets show that the segmentation accuracy of the proposed model for dripping gasoline is 90.34%,and the average detection time of each image is 0.23 s.Compared with Unet

关 键 词：汽油管道泄露 高光谱图像 目标检测 深度学习 Unet网络 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]