基于改进DeepLabV3+的引导式道路提取方法及在震源点位优化中的应用  

Guided Road Extraction Method Based on Improved DeepLabV3+and Its Application in Optimization of Source Positions

作　　者：曹凯奇 张凌浩 徐虹[1] 吴蔚 文武[1] 周航 CAO Kaiqi;ZHANG Linghao;XU Hong;WU Wei;WEN Wu;ZHOU Hang(School of Computer Science,Chengdu University of Information Technology,Chengdu,Sichuan 610225,China;Electric Power Research Institute,State Grid Sichuan Electric Power Company,Chengdu,Sichuan 610094,China;Collection Technology Center,Bureau of Geophysical Prospecting INC of China National Petroleum Corporation,Zhuozhou,Hebei 072750,China)

机构地区：[1]成都信息工程大学计算机学院,四川成都610225 [2]国网四川省电力公司电力科学研究院,四川成都610094 [3]中国石油集团东方地球物理勘探有限责任公司采集技术中心,河北涿州072750

出　　处：《西安石油大学学报（自然科学版）》2024年第2期128-142,共15页Journal of Xi’an Shiyou University（Natural Science Edition）

基　　金：国家自然科学基金面上项目“基于频变信息的流体识别及流体可动性预测”(41774142);四川省重点研发项目“工业互联网安全与智能管理平台关键技术研究与应用”(2023YFG0112);四川省自然科学基金资助项目“基于超分辨感知方法的密集神经图像分割”(2022NSFSC0964)。

摘　　要：为解决自动识别方法在道路提取时存在漏提、错提现象,提出一种引导式道路提取方法提高修正效率。在DeepLabV3+原有输入通道(3通道)的基础上添加额外输入通道(第4通道),将道路的4个极点转化为二维高斯热图后作为额外通道输入网络,网络以极点作为引导信号,使网络适用于引导式道路提取任务;设计并行多分支模块,提取上下文信息,增强网络特征提取能力;融合类均衡二值交叉熵和骰子系数组成新的复合损失函数进行训练缓解正负样本不均衡问题。在公共Deepglobe数据集和西南某区域三维实际数据集上对本文网络进行验证,在Deepglobe上的像素精确度PA、交并比IOU、F1分数分别达到82.29%、68.81%和81.52%;在西南某区域三维数据集上PA、IOU、F1分别达到89.05%、81.01%和89.51%。实际应用表明:该方法能够有效提高道路识别精度,道路符合率达到85%以上,为后续震源点布设提供准确的信息。In order to solve the problem of missing or incorrect extraction in automatic recognition method,a guided road information extraction method is proposed to improve the correction efficiency.This method adds an additional input channel(4th channel)in three original input channels of DeepLabV3+,converting the four poles of the road into a two-dimensional Gaussian heat map as additional channel input to the network.Taking the poles as guidance signals makes the network suitable for guided road information extraction tasks.The parallel multi-branch modules are designed to extract contextual information and enhance the feature extraction capabilitiy of the network.A new composite loss function is formed by fusing the class equilibrium binary cross entropy and dice coefficient to train and alleviate the imbalance of positive and negative samples.The network in this paper was verified on the public Deepglobe dataset and a actual 3D dataset in southwest China.PA,IOU and F1 on the public Deepglobe dataset reached 82.29%,68.81%and 81.52%respectively;On the 3D dataset,PA,IOU,and F1 reached 89.05%,81.01%and 89.51%respectively.Practical applications show that this method can effectively improve the accuracy of road recognition,with a compliance rate of over 85%with the road,providing accurate information for the subsequent deployment of seismic source points.

关 键 词：道路拾取 深度学习 DeepLabV3+ 震源点布设 

分 类 号：TE19[石油与天然气工程—油气勘探]