en
×

分享给微信好友或者朋友圈

使用微信“扫一扫”功能。
作者简介:

于会臻(1981—),男,山东桓台人,副研究员,博士,从事综合地球物理勘探技术研究。E-mail:yhzabc@163.com。

中图分类号:TE319

文献标识码:A

文章编号:1009-9603(2022)06-0058-09

DOI:10.13673/j.cnki.cn37-1359/te.202111016

参考文献 1
SMITH D A.Sealing and nonsealing faults in Louisiana Gulf Coast salt basin[J].AAPG Bulletin,1980,64(2):145-172.
参考文献 2
HOOPER E C D.Fluid migration along growth faults in compact⁃ ing sediments[J].Journal of Petroleum Geology,1991,14(S1):161-180.
参考文献 3
李强,田晓平,何京,等.断层封闭性定量表征及对油气富集的控制作用[J].西南石油大学学报:自然科学版,2018,40(4):40-50.LI Qiang,TIAN Xiaoping,HE Jing,et al.Quantitative character⁃ ization of fault sealing and its control on hydrocarbon accumula⁃ tion[J].Journal of Southwest Petroleum University:Science & Technology Edition,2018,40(4):40-50.
参考文献 4
MARFURT K J,KIRLIN R L,FARMER S L,et al.3-D seismic at⁃ tributes using a semblance-based coherency algorithm[J].Geo⁃ physics,1998,63(4):1 150-1 165.
参考文献 5
陆文凯,张善文,肖焕钦.基于相干滤波的相干体图像增强[J].天然气工业,2006,26(5):37-39.LU Wenkai,ZHANG Shanwen,XIAO Huanqin.Image enhance⁃ ment of coherence cube based on coherent filtering[J].Natural Gas Industry,2006,26(5):37-39.
参考文献 6
马承杰.多尺度边缘检测技术在断层识别及裂缝发育带预测中的应用——以车排子地区排691井区为例[J].油气地质与采收率,2021,28(2):85-90.MA Chengjie.Application of multi-scale edge detection technolo⁃ gy to fault recognition and fracture zone prediction:A case study of Block Well P691,Chepaizi area[J].Petroleum Geology and Re⁃ covery Efficiency,2021,28(2):85-90.
参考文献 7
RANDEN T,PEDERSEN S I,SØNNELAND L.Automatic extrac⁃ tion of fault surfaces from three-dimensional seismic data[C].San Antonio:SEG International Exposition and Annual Meeting,2001:551-554.
参考文献 8
ROBERTS A.Curvature attributes and their application to 3D in⁃ terpreted horizons[J].First Break,2001,19(2):85-100.
参考文献 9
DI H B,GAO D.Efficient volumetric extraction of most positive/negative curvature and flexure for fracture characterization from 3D seismic data[J].Geophysical Prospecting,2016,64(6):1 454-1 468.
参考文献 10
HALE D.Methods to compute fault images,extract fault surfaces,and estimate fault throws from 3D seismic images[J].Geophysics,2013,78(2):O33-O43.
参考文献 11
PEDERSEN S I,RANDEN T,SONNELAND L,et al.Automatic fault extraction using artificial ants[C].Salt Lake City:72nd An⁃ nual International Meeting,SEG,Expanded Abstracts,2002.
参考文献 12
PEDERSEN S I,SKOV T,HETLELID A,et al.New paradigm of fault interpretation[C].Dallas:73rd Annual International Meet⁃ ing,SEG,Expanded Abstracts,2003.
参考文献 13
WU Xinming,ZHU Zhihui.Methods to enhance seismic faults and construct fault surfaces[J].Computers & Geosciences,2017,107:37-48.
参考文献 14
WU Xinming,FOMEL S.Automatic fault interpretation with opti⁃ mal surface voting[J].Geophysics,2018,83(5):O67-O82.
参考文献 15
WU Xinming,LIANG Luming,SHI Yunzhi,et al.FaultSeg3D:Us⁃ ing synthetic datasets to train an end-to-end convolutional neural network for 3D seismic fault segmentation[J].Geophysics,2019,84(3):35-45.
参考文献 16
杨午阳,杨佳润,陈双全,等.基于U-Net深度学习网络的地震数据断层检测[J].石油地球物理勘探,2021,56(4):688-697.YANG Wuyang,YANG Jiarun,CHEN Shuangquan,et al.Seismic data fault detection based on U-Net deep learning network[J].Oil Geophysical Prospecting,2021,56(4):688-697.
参考文献 17
CHEHRAZI A,RAHIMPOUR-BONAB H,REZAEE M R.Seis⁃ mic data conditioning and neural network-based attribute selec⁃ tion for enhanced fault detection[J].Petroleum Geoscience,2013,19(2):169-183.
参考文献 18
ZHANG C,FROGNER C,ARAYA-POLO M,et al.Machinelearning based automated fault detection in seismic traces[C].Amsterdam:76th EAGE Conference and Exbibition,2014:1-5.
参考文献 19
HUANG L,DONG X S,CLEE T E.A scalable deep learning plat⁃ form for identifying geologic features from seismic attributes[J].The Leading Edge,2017,36(3):249-256.
参考文献 20
XIONG W,JI X,MA Y,et al.Seismic fault detection with convolu⁃ tional neural network[J].Geophysics,2018,83(5):O97-O103.
参考文献 21
MA Yide,LIU Qing,QIAN Zhibai.Automated image segmentation using improved PCNN model based on cross-entropy[C].Hong Kong:Proceedings of 2004 International Symposium on Intelligent Multimedia,Video and Speech Processing,2004:743-746.
参考文献 22
LIN T Y,GOYAL P,GIRSHICK R,et al.Focal loss for dense ob⁃ ject detection[C].Venice:IEEE Transactions on Pattern Analysis and Machine Intelligence,2017:2 999-3 007.
参考文献 23
SUDRE C H,LI Wenqi,VERCAUTEREN Tom,et al.Generalised Dice overlap as a deep learning loss function for highly unbal⁃ anced segmentations[C].Quebec City:International Workshop on Deep Learning in Medical Image Analysis International Workshop on Multimodal Learning for Clinical Decision Support,2017:240-248.
参考文献 24
JADON S.A survey of loss functions for semantic segmentation [C].Viña del Mar:IEEE Conference on Computational Intelli⁃ gence in Bioinformatics and Computational Biology,2020:1-7.
参考文献 25
CHEN L C,PAPANDREOU G,KOKKINOS I,et al.Semantic im⁃ age segmentation with deep convolutional nets and fully connect⁃ ed CRFs[J].Computer Science,2014,(4):357-361.
参考文献 26
CHEN L C,PAPANDREOU G,KOKKINOS I,et al.DeepLab:Se⁃ mantic image segmentation with deep convolutional nets,atrous convolution,and fully connected CRFs[J].IEEE Transactions on Pattern Analysis and Machine Intelligence,2018,40(4):834-848.
参考文献 27
CHEN L C,ZHU Yukun,PAPANDREOU G,et al.Encoder-de⁃ coder with atrous separable convolution for semantic image seg⁃ mentation[C].Munich:Proceedings of the European Conference on Computer Vision,2018:801-818.
参考文献 28
ZHAO H S,SHI J P,QI X J,et al.Pyramid scene parsing network [C].Hawaii:IEEE Conference on Computer Vision and Pattern Recognition,2017:6 230-6 239.
参考文献 29
SHELHAMER E,LONG J,DARRELL T.Fully convolutional net⁃ works for semantic segmentation[J].IEEE Transactions on Pattern Analysis and Machine Intelligence,2017,39(4):640-651.
参考文献 30
RONNEBERGER O,FISCHER P,BROX T.U-Net:Convolutional networks for biomedical image segmentation[C].Munich:Medical Image Computing and Computer Assisted Intervention,2015:234-241.
参考文献 31
秦飞巍,沈希乐,彭勇,等.无人驾驶中的场景实时语义分割方法[J].计算机辅助设计与图形学学报,2021,33(7):1 026-1 037.QIN Feiwei,SHEN Xiyue,Peng Yong,et al.A real-time semantic segmentation approach for autonomous driving scenes[J].Journal of Computer-Aided Design & Computer Graphics,2021,33(7):1 026-1 037.
参考文献 32
胡嵽,冯子亮.基于深度学习的轻量级道路图像语义分割算法 [J].计算机应用,2021,41(5):1 326-1 331.HU Die,FENG Ziliang.Light-weight road image semantic seg⁃ mentation algorithm based on deep learning[J].Journal of Com⁃ puter Applications,2021,41(5):1 326-1 331.
参考文献 33
方超伟,李雪,李钟毓,等.基于双模型交互学习的半监督医学图像分割[J].自动化学报,2020,46(13):1-15.FANG Chaowei,LI Xue,LI Zhongyu,et al.Interactive dual-mod⁃ el learning for semi-supervised medical image segmentation[J].Acta Automatica Sinica,2020,46(13):1-15.
参考文献 34
徐世杰,杜煜,鹿鑫,等.基于ENet的轻量级语义分割算法研究 [J].计算机工程与科学,2021,43(8):1 454-1 460.XU Shijie,DU Yu,LU Xin,et al.A lightweight semantic segmen⁃ tation algorithm based on ENet[J].Computer Engineering & Sci⁃ ence,2021,43(8):1 454-1 460.
参考文献 35
FAN Heng,LING Haibin.SANet:Structure-aware network for vi⁃ sual tracking[C].Hawaii:IEEE Conference on Computer Vision and Pattern Recognition Workshops,2017:2 217-2 224.
参考文献 36
HUANG Zilong,WANG Xinggang,HUANG Lichao,et al.CCNet:Criss-cross attention for semantic segmentation[C].Seoul:IEEE/CVF International Conference on Computer Vision,2019:603-612.
参考文献 37
FU Jun,LIU Jing,TIAN Haijie,et al.Dual attention network for scene segmentation[C].Long Beach:IEEE/CVF Conference on Computer Vision and Pattern Recognition,2020:3 141-3 149.
参考文献 38
HUANG Huimin,LIN Lanfen,TONG Ruofeng,et al.UNet3+:A full-scale connected UNet for medical image segmentation[C].Barcelona:IEEE International Conference on Acoustics,Speech and Signal Processing,2020:1 055-1 059.
参考文献 39
ZHOU Zongwei,SIDDIQUEE M M R,TAJBAKHSH N,et al.UN⁃ et++:A nested U-Net architecture for medical image segmentation [C].Berlin:Deep Learning in Medical Image Analysis and Multi⁃ modal Learning for Clinical Decision Support,2018:3-11.
参考文献 40
VASWANI A,RAMACHANDRAN P,SRINIVAS A,et al.Scaling local self-attention for parameter efficient visual backbones[C].Nashville:IEEE/CVF Conference on Computer Vision and Pattern Recognition,2021:12 889-12 899.
参考文献 41
KINGMA D P,BA J L.Adam:A method for stochastic optimization [C].San Diego:3rd International Conference on Learning Repre⁃ sentations,2015:1-15.
目录contents

    摘要

    断层解释是油气勘探开发的基础工作之一。近年来,深度学习凭借其强大的数据分析能力,为精细刻画断层空间展布提供了新的技术工具,其应用前提在于如何获取大量可靠的样本数据。与目前最流行的断层正演样本数据相比,实际资料的断层专家解释样本数据存在一定的主观性,通常仅关注目标区域,断层样本标签并不完备。改进后的不完备断层样本标签损失函数将断层分析聚焦于已解释区域,提高了断层专家解释样本数据的可用性;通过对断层专家解释结果的样本数据增强处理,扩充了有效样本数据量;设计构建了联合专家解释样本及正演样本的三维断层自动识别网络结构,并在其中引入自注意力机制,提升了三维断层自动识别网络模型的泛化能力和空间特征分析能力。模型试验及实际应用测试结果表明,三维断层自动识别方法可融入实际断层发育特征,识别结果更符合地质认识,且精度也得到有效提升,从而验证了其可靠性及实用性。

    Abstract

    Fault interpretation is one of the basic works of oil & gas exploration and development. In recent years,with its powerful data analysis ability,deep learning has provided a new technical tool for characterizing the spatial distribution of faults in detail. Its application relies on how to obtain a large number of reliable sample data. Compared with the most popu- lar fault data from forward modeling samples at present,fault data from expert interpretation samples of actual data are not only subjective but also focus on target areas,with incomplete labels of fault samples. The improved loss function of incom- plete labels of fault samples emphasizes interpreted areas in fault analysis,which improves the availability of fault data from expert interpretation samples. Through the enhancement processing of data from expert fault interpretation samples, the amount of effective sample data is increased. In addition,this paper designs and constructs a network structure of an au- tomatic 3D fault identification method driven by data from expert interpretation samples and forward modeling samples,and it introduces a self-attention mechanism to improve the generalization ability and spatial feature analysis ability of the auto- matic 3D fault identification network model. Model tests and practical application show that the proposed automatic 3D fault identification method can analyze actual features of the fault development. The identification results are more in line with geological conditions and the accuracy is effectively improved,which verifies the reliability and practicality of this method.

  • 断层在油气运移和聚集过程中发挥着重要的作用[1-3],但精细的断层识别结果需要开展大量繁琐的三维地震解释工作。为有效提升断层识别效率及精度,相干[4-5]、边缘检测[6]、方差[7]、曲率[8-9] 和断层似然性[10] 等常规断层识别技术得到不断发展,研究思路主要是通过人工设计的算法及流程凸显同相轴的不连续性,实现断层特征的增强。然而受观测数据和处理噪声等非地质因素的影响,识别结果中包含较多干扰信息,需利用蚂蚁追踪[11-12]、平滑滤波[13]、最优面投票[14] 等算法来进行后续处理,在算法应用过程中参数选择困难,存在较强主观性,效率和可靠性有待进一步提升。近年来,随着大数据及人工智能技术的发展,基于深度学习的地震自动解释方法得到广泛关注。在基于深度学习的断层自动识别方法中,根据样本数据来源将其大致分为 2 类:第一类是基于正演模拟方法构建样本数据[14-15];第二类是基于实际资料构建样本数据[16-20]。 2 类方法各有优势及不足,第一类方法的优势在于可快速形成深度学习所需的大量样本数据,但问题在于理论模型无法完全反映复杂的真实断层分布特征及地震反射特征,复杂断层的识别精度有待进一步提升;第二类方法虽然利用实际资料进行模型训练,但构建大量可靠的三维断层专家解释识别样本库较困难,数据不足和复杂断层识别的主观性会严重影响断层识别网络模型的泛化能力。如何充分融合2类断层样本数据的优势还需深入研究。为此,笔者提出了一种联合专家解释样本及正演样本的三维断层自动识别方法,主要包含3个关键内容:一是不完备断层样本标签损失函数的改进,为降低未解释区域断层样本标签不明确对断层识别网络模型参数优化带来的干扰,将断层识别损失函数计算的重点集中于已解释断层附近的位置区域,即断层解释聚焦区域;二是二维断层专家解释样本数据的增强处理,通过二维样本数据维度扩充、三维断面扭曲、三维空间旋转等方法对二维断层专家解释样本数据进行三维增强处理,以增加专家解释样本的多样性;三是三维断层自动识别网络结构的设计,引入自注意力机制,构建了联合专家解释样本及正演样本的三维断层自动识别网络结构,断层识别精度和可靠性得到进一步提升。

  • 1 方法原理

  • 1.1 不完备断层专家解释样本标签损失函数改进

  • 断层识别深度神经网络正向传播过程可表达为:

  • f(WS)F~S=Sr,SfF~=F~r,F~f
    (1)
  • 为保证地震数据经过神经网络得到的断层识别输出结果与断层样本标签最为匹配,期望得到的深度神经网络权重参数 W 应满足预测结果和已知标签之间的目标函数最小,其表达式为:

  • J=argminwL(F,F~)F=Fr,Ff
    (2)
  • 假设在断层的采样点位置标记为 1 类,在非断层的采样点位置标记为 0 类,则在三维地震数据体中识别断层的问题可类比为二分类的语义分割问题。常用的语义分割损失函数包括交叉熵损失 (Cross-Entropy)[21]、焦点损失函数(Focal Loss)[22]、骰子损失函数(Dice Loss)[23] 等。由于断层在三维空间中的数量远远少于非断层的数量,样本类别存在不平衡的问题,因此选用了基于Dice Loss改进后的 Log Cosh Dice Loss损失函数[24],其表达式为:

  • L^=1-2FF~+1F+F~+1L'=logeL^+e-L^2
    (3)
  • 受勘探认识程度和工作量等客观因素的影响,往往只对三维地震工区的重点区域进行精细的断层解释,未解释区域缺少断层专家解释结果,且样本标签存在不完备性问题。为此,常规损失函数已不适用于此类断层样本标签,而应着重关注已解释断层附近的位置区域,即断层专家解释聚焦区域,以增强断层自动识别网络模型的可靠性。

  • 基于上述讨论,断层专家解释聚焦区域的生成方式为:

  • B=g*Fr
    (4)
  • B的尺寸与地震剖面和断层专家解释样本标签 Fr的尺寸相同。在此基础上,构建断层专家解释聚焦区域掩码权重M(尺寸与B相同),其表达式为:

  • (5)
  • 则改进后的不完备断层专家解释样本标签损失函数表达式为:

  • Lnew =minW L'MF~r,MFr+minW L'(1-M)F~r,(1-M)Fr
    (6)
  • 1.2 二维断层专家解释样本数据增强处理

  • 实际资料的地震解释结果多为二维数据,为将其与三维正演样本数据进行有机融合,需要对其进行维度扩充处理,并在此基础上,利用三维断面扭曲、三维空间旋转等方法对样本数据进行增强处理,以提供更丰富的断层专家解释样本数据,从而进一步增加三维断层识别网络模型的泛化能力。具体步骤如下:

  • 第 1 步:二维断点专家解释结果的空间插值。在三维直角坐标系(XYT)下,设二维实际地震剖面Sr0RNt×Nx,其XT方向的原始位置向量分别为 x0t0,对专家解释的断点进行分段插值得到连续的断层,如图1a红色线所示。红色断层线附近的采样点位置赋值为1,表示为已解释断层;其他采样点位置赋值为0,表示非断层,其中也包含未解释区域的断层,则可得到与实际地震剖面相对应的二维断层专家解释样本标签Fr0

  • 第 2 步:二维样本数据维度扩充。将二维实际地震剖面 Sr0和二维断层专家解释样本标签 Fr0沿着垂直于二维剖面 Y方向进行复制扩充(图1b),得到维度扩充后的三维实际地震数据体 Sr1和三维断层专家解释样本标签数据体Fr1尺寸变为Nt × Nx × Ny,其XYT方向的原始位置向量分别为x0y0t0

  • 第 3 步:三维断面扭曲。为进一步丰富断层样本数据的样式,将维度扩充后得到的三维实际地震数据体 Sr1和三维断层专家解释样本标签数据体 Fr1 沿着X方向进行任意扭曲。采用多个正弦信号叠加的形式模拟扭曲的断面形态(图1c),其表达式为:

  • x1=k=1K aksinwky0+ψk
    (7)
  • 按照三维断面扭曲后的位置向量 x1y0t0对三维实际地震样本数据体 Sr1和三维断层专家解释样本标签数据体Fr1进行插值,获得断面扭曲后的三维实际地震数据体 Sr2和三维断层专家解释样本标签数据体Fr2

  • 第 4 步:三维空间旋转。对断面扭曲后的三维实际地震数据体 Sr2和三维断层专家解释样本标签数据体Fr2沿着不同方位角进行旋转(图1d),其表达式为:

  • x2y1=expθ0-110x1y0
    (8)
  • 按照三维空间旋转后的位置向量 x2y1t0对三维实际地震样本数据体 Sr2和三维断层专家解释样本标签数据体Fr2进行插值,获得空间旋转后的三维实际地震数据体 Sr3和三维断层专家解释样本标签数据体Fr3

  • 重复第 3 和第 4 步,随机设置不同的扭曲和方位旋转参数,可获得大量不同弯曲形态和旋转角度的三维实际地震数据体和三维断层专家解释样本标签数据体的增强处理结果,增加实际资料断层专家解释样本数据的多样性。

  • 图1 断层专家解释样本数据增强图

  • Fig.1 Fault data enhancement of expert interpretation samples

  • 1.3 三维断层自动识别网络结构设计

  • 断层识别可类比为图像语义分割问题。在该领域研究中,DeepLab[25-27],PSPNet[28],FCN[29]和 U-Net[30]等网络结构已在自动驾驶[31-32]、医学图像分析[33-34] 等语义分割应用场景中取得了理想的应用效果。同时,研究者在上述网络结构的基础上,不断完善中间层特征提取能力,引入了在自然语言处理领域取得应用突破的自注意力机制,提出了 SANet[35],CCNet[36]和 DANet[37]等网络结构,语义分割效果得到进一步提升。借鉴上述研究思路,构建了联合专家解释样本及正演样本的三维断层自动识别网络结构(图2)。

  • 基本网络结构借鉴UNet3+ [38],主要包括编码和解码 2 个过程(图2)。编码过程包含 4 个特征提取块(E1E2E3E4),从小尺度到大尺度逐步提取地震反射特征;解码过程包含 3 个特征恢复块(D1D2D3),分别与前3个特征提取块(E1E2E3)一一对应,分别恢复不同尺度特征的断层识别结果。与U-Net 和 UNet++[39] 相比,UNet3+重新设计了跳跃连接,解码过程可直接融合来自不同尺度的特征提取块和上层特征恢复块的特征,既能捕获全尺度下的细粒度语义和粗粒度语义,又能减少训练参数。同时,对每个特征恢复块输出结果进行深度监督,即在每一个特征恢复块模块的最后一层被送入 1个 3×3×3 的卷积层,并通过 1 个双线性上采样层得到样本标签一样大小的特征图,再加入 1 个 Sigmoid 函数层,之后利用(6)式分别计算不同尺度断层识别结果与样本标签间的匹配程度。

  • 构建的网络结构不仅需要学习样本数据与标签之间的关系,还应能分析断面上不同断点及其地震反射特征的空间相关关系。通常采用增加卷积神经网络的层数或改变卷积核尺寸的方法来增大视野、增强多尺度分析能力,但网络层数过多会使权重参数的数量及求解难度增加,故盲目扩大卷积核尺寸会导致细节信息的丢失。自注意力机制可动态计算中间层的空间结构特征,具有比卷积层更大的视野。为此,在每个特征提取块和特征恢复块都添加了具有更强数据空间特征分析能力的 Halo⁃ Net[40] 注意力层,即每个特征提取块改为一个3×3×3 的卷积层、HaloNet 注意力层、ReLU 激活函数和 2× 2×2 的最大池化层的连接,每个特征恢复块改为一个3×3×3的转置卷积层、HaloNet注意力层、ReLU激活函数的连接。HaloNet采用分块自注意力机制,既保证大规模样本数据计算效率的提升,也有助于获得更高精度的三维断层自动识别网络模型。

  • 2 技术流程

  • 综上所述,制定了联合专家解释样本及正演样本的三维断层自动识别方法的技术流程(图3),具体步骤包括:①二维断层专家解释样本数据标准化处理。针对不同工区的二维实际地震数据及断层专家解释样本标签数据,沿着垂直时间和水平地震道方向分别进行重采样或插值处理,将数据的时间间隔和道间距进行统一,保证训练数据的一致性。 ②二维断层专家解释样本数据增强处理。对一致性处理后的地震数据和断层专家解释样本标签进行维度扩充及随机的断面扭曲和方位三维空间旋转,形成大量的三维断层专家解释样本数据。③三维断层正演样本数据构建。根据文献[15]方法,针对各类断层发育模式,设置地层褶皱幅度、断层倾角、方位角、子波峰值频率和噪声等参数,形成大量三维断层正演样本数据。④三维断层自动识别网络结构构建。建立联合专家解释样本及正演样本的深度学习断层自动识别网络。⑤三维断层自动识别网络参数设置及优化。整理增强后的三维断层专家解释样本数据和三维断层正演样本数据,按照K折交叉验证流程形成训练集和验证集。设置最大迭代轮次、学习率等网络训练关键参数,在样本数据的驱动下,利用 Adam[41]优化算法进行迭代求解,得到三维断层自动识别网络模型。⑥验证集评价分析及获得最终三维断层自动识别网络模型。将训练得到的三维断层自动识别网络模型在验证集上进行测试评价,当满足评价目标时,将该三维断层自动识别网络模型作为最终模型。否则,反复修正网络参数,直至满足评价要求。⑦实际应用。将最终获得的三维断层自动识别网络模型应用于实际工区,实现高效、高精度的三维断层自动识别。

  • 图2 联合专家解释样本及正演样本的三维断层自动识别网络结构

  • Fig.2 Network structure of automatic 3D fault identification method driven by data from expert interpretation samples and forward modeling samples

  • 图3 联合专家解释样本及正演样本的三维断层自动识别方法技术流程

  • Fig.3 Technical flowchart of automatic 3D fault identification method driven by data from expert interpretation samples and forward modeling samples

  • 3 模型验证

  • 3.1 数据预处理

  • 通过三维断层正演模型来验证本文方法的可靠性。构建了200对尺寸为256×512×256(nx×ny×nt)的三维断层模型及地震数据作为模型实验样本集。将该样本集分为 3 份,其中 40% 作为待解释的三维地震数据,40% 作为三维断层正演样本数据,另外 20%作为验证集。

  • 对于待解释的三维地震数据,进行随机剖面断层解释,断层解释数量为所有断层数量的 75%。利用所提的样本数据增强处理方法得到断层专家解释样本数据集合,与其他三维断层正演样本数据进行联合,作为最终使用的训练集。采用图2 网络结构,利用 Adam 算法进行参数优化,学习率设置为 0.000 1,迭代轮次为 100。通过不断迭代训练得到三维断层自动识别网络模型。

  • 3.2 效果分析

  • 图4a为验证集中的任意二维地震剖面数据,图4b 为其对应的断层专家解释样本标签数据。对比常规损失函数的断层自动识别结果(图4c)和损失函数改进后的断层自动识别结果(图4d)可以看出,由于训练集断层样本标签的不完备性,常规损失函数无法准确评估地震响应特征是否对应断层,反向传播时影响了权重参数求解的可靠性,导致部分断层没有得到有效识别。而利用改进后的损失函数,在已解释断层的聚焦区域内进行地震反射特征与断层相关关系分析,网络训练的目标更加明确,断层自动识别与真实标签的吻合度显著提高,验证集的模型测试正确率为 96% 左右。同时,在优化方法、学习率、迭代轮次等超参数固定的前提下,相比 U-Net,UNet++,UNet3+等网络结构,所提出的网络结构更好地提升了地震空间特征分析能力,断层识别正确率可提升3%~4%。

  • 4 应用实例

  • 选取胜利油田东部某工区开展应用测试。该试验区位于东营凹陷东北部,为一复杂断块油气藏。断层样本数据包括收集整理的相邻工区共 80 个地震剖面断层专家解释结果和正演模拟构建的 400对 256×256×256三维断层样本数据。网络训练采用Adam算法进行优化,学习率为0.000 1,迭代轮次为 100,通过反复迭代优化得到最终断层自动识别网络模型。

  • 对比地震断层属性与深度学习断层识别方法,从图5a可以看出,本征相干属性检测到地震反射不连续性,但断层识别结果噪声干扰强,垂向上连续性差,不利于开展后续的解释工作。图5b是常规深度学习的断层自动识别结果,仅采用了正演样本数据和常规 UNet3+网络进行断层网络模型的训练和自动识别。相比图5a来说,新方法的断层自动识别剖面(图5c)和三维显示(图5d)中三维断层自动识别结果都具有更好的空间连续性、抗噪性,究其原因是通过多层次的特征提取及恢复,同时考虑了小时窗内的断层细节变化和大时窗内的断层结构变化,从而提升了断层识别结果的可解释性。

  • 图4 损失函数改进前后的断层自动识别结果对比

  • Fig.4 Comparison of automatic fault identification results before and after improvement of loss function

  • 图5 不同方法下断层自动识别结果对比

  • Fig.5 Comparison of automatic fault identification results by different methods

  • 对比图5b和图5c可知,新方法在融合了基于专家经验的断层解释样本数据之后,断层识别结果更加符合地质经验认识。例如,对于地震剖面右下方的低角度断层,常规深度学习的断层自动识别结果 (图5b)未能准确预测,而新方法的断层自动识别结果(图5c,5d)有效地识别出了该断层,与已知地质认识更加吻合,可为后续断面组合分析提供更为可靠的参考依据,也验证了该方法的实用性。

  • 5 结论

  • 提出了一种联合专家解释样本及正演样本的三维断层自动识别的新方法。断层样本数据的数据量、多样性和可靠性是影响断层自动识别应用效果的重要因素,为更好地发挥断层专家解释样本数据与断层正演样本数据的各自优势,在不完备断层样本标签损失函数、二维断层专家解释样本数据增强处理、联合专家解释样本及正演样本的三维断层自动识别网络结构等 3 个方面进行了改进和创新。相比现有的深度学习断层自动识别方法,笔者所提的新方法既能充分融合丰富的实际地震数据和断层发育特征,还能加入大量的三维断层正演样本数据进行共同训练,在保证识别效率的前提下,三维断层自动识别网络模型的泛化能力和精度都得到进一步提升。

  • 新方法的研究思路不仅适用于断层自动识别,也可应用于岩相岩性识别、层位解释等其他智能地震解释工作。在下步工作中,将根据不同构造样式、岩性变化等地质情况,优化样本增强和正演模拟流程,构建规模更大、样式更丰富的专家解释样本及正演样本数据库,并开展地震资料自动优化、已有构造解释成果约束、低序级断层自动识别、断面自动组合等方法的深入研究,更加系统地考虑断层识别问题,从而实现高精度三维断裂系统的快速构建。

  • 符号解释

  • B——断层专家解释聚焦区域;

  • D1D2D3——三维断层自动识别网络的 3 个特征恢复块;

  • E1E2E3E4——三维断层自动识别网络的 4 个特征提取块;

  • f(·)——非线性激活函数;

  • F——断层样本标签;

  • Ff ——正演模型的断层样本标签;

  • Fr ——断层专家解释样本标签;

  • Fr0——二维断层专家解释样本标签;

  • Fr1——维度扩充后的三维断层专家解释样本标签数据体;

  • Fr2——断面扭曲后的三维断层专家解释样本标签数据体;

  • Fr3——空间旋转后的三维断层专家解释样本标签数据体;

  • F~——地震数据经过神经网络得到的断层识别输出结果;

  • F~f——地震正演模拟数据经过神经网络得到的断层识别输出结果;

  • F~r——实际地震数据经过神经网络得到的断层识别输出结果;

  • g——全部元素为1的二维卷积核;

  • J——目标函数;

  • k——模拟断层弯曲形态所需正弦弯曲曲线编号,取值为1~K

  • K——模拟断层弯曲形态所需正弦弯曲曲线的总个数;

  • L——神经网络损失函数;

  • L^——骰子损失函数;

  • L′——Log Cosh Dice Loss损失函数;

  • Lnew——改进后的不完备断层专家解释样本标签损失函数;

  • M——断层专家解释聚焦区域掩码权重;

  • nt ——时窗垂直时间采样点个数;

  • nx——时窗水平方向道的个数;

  • ny——时窗水平方向线的个数;

  • Nt ——地震数据的双程旅行时间T方向维度;

  • Nx——地震数据X方向维度;

  • Ny——地震数据Y方向维度;

  • S——地震数据;

  • Sf ——断层模型地震正演模拟数据;

  • Sr ——实际地震数据;

  • Sr0——二维实际地震剖面;

  • Sr1——维度扩充后的三维实际地震数据体;

  • Sr2——断面扭曲后的三维实际地震数据体;

  • Sr3——空间旋转后的三维实际地震数据体;

  • t0——二维剖面T方向原始位置向量;

  • wk——第k个正弦弯曲曲线的角频率;

  • W——深度神经网络权重参数;

  • x0——二维剖面X方向原始位置向量;

  • x1——三维断面扭曲后的X方向位置向量;

  • x2——三维空间旋转后的X方向位置向量;

  • y0——维度扩充后的三维数据Y方向原始位置向量;

  • y1——三维空间旋转后的Y方向位置向量;

  • ak——第k个正弦弯曲曲线的振幅;

  • ψk——第k个正弦弯曲曲线的初相;

  • θ——旋转的方位角,(°)。

  • 参考文献

    • [1] SMITH D A.Sealing and nonsealing faults in Louisiana Gulf Coast salt basin[J].AAPG Bulletin,1980,64(2):145-172.

    • [2] HOOPER E C D.Fluid migration along growth faults in compact⁃ ing sediments[J].Journal of Petroleum Geology,1991,14(S1):161-180.

    • [3] 李强,田晓平,何京,等.断层封闭性定量表征及对油气富集的控制作用[J].西南石油大学学报:自然科学版,2018,40(4):40-50.LI Qiang,TIAN Xiaoping,HE Jing,et al.Quantitative character⁃ ization of fault sealing and its control on hydrocarbon accumula⁃ tion[J].Journal of Southwest Petroleum University:Science & Technology Edition,2018,40(4):40-50.

    • [4] MARFURT K J,KIRLIN R L,FARMER S L,et al.3-D seismic at⁃ tributes using a semblance-based coherency algorithm[J].Geo⁃ physics,1998,63(4):1 150-1 165.

    • [5] 陆文凯,张善文,肖焕钦.基于相干滤波的相干体图像增强[J].天然气工业,2006,26(5):37-39.LU Wenkai,ZHANG Shanwen,XIAO Huanqin.Image enhance⁃ ment of coherence cube based on coherent filtering[J].Natural Gas Industry,2006,26(5):37-39.

    • [6] 马承杰.多尺度边缘检测技术在断层识别及裂缝发育带预测中的应用——以车排子地区排691井区为例[J].油气地质与采收率,2021,28(2):85-90.MA Chengjie.Application of multi-scale edge detection technolo⁃ gy to fault recognition and fracture zone prediction:A case study of Block Well P691,Chepaizi area[J].Petroleum Geology and Re⁃ covery Efficiency,2021,28(2):85-90.

    • [7] RANDEN T,PEDERSEN S I,SØNNELAND L.Automatic extrac⁃ tion of fault surfaces from three-dimensional seismic data[C].San Antonio:SEG International Exposition and Annual Meeting,2001:551-554.

    • [8] ROBERTS A.Curvature attributes and their application to 3D in⁃ terpreted horizons[J].First Break,2001,19(2):85-100.

    • [9] DI H B,GAO D.Efficient volumetric extraction of most positive/negative curvature and flexure for fracture characterization from 3D seismic data[J].Geophysical Prospecting,2016,64(6):1 454-1 468.

    • [10] HALE D.Methods to compute fault images,extract fault surfaces,and estimate fault throws from 3D seismic images[J].Geophysics,2013,78(2):O33-O43.

    • [11] PEDERSEN S I,RANDEN T,SONNELAND L,et al.Automatic fault extraction using artificial ants[C].Salt Lake City:72nd An⁃ nual International Meeting,SEG,Expanded Abstracts,2002.

    • [12] PEDERSEN S I,SKOV T,HETLELID A,et al.New paradigm of fault interpretation[C].Dallas:73rd Annual International Meet⁃ ing,SEG,Expanded Abstracts,2003.

    • [13] WU Xinming,ZHU Zhihui.Methods to enhance seismic faults and construct fault surfaces[J].Computers & Geosciences,2017,107:37-48.

    • [14] WU Xinming,FOMEL S.Automatic fault interpretation with opti⁃ mal surface voting[J].Geophysics,2018,83(5):O67-O82.

    • [15] WU Xinming,LIANG Luming,SHI Yunzhi,et al.FaultSeg3D:Us⁃ ing synthetic datasets to train an end-to-end convolutional neural network for 3D seismic fault segmentation[J].Geophysics,2019,84(3):35-45.

    • [16] 杨午阳,杨佳润,陈双全,等.基于U-Net深度学习网络的地震数据断层检测[J].石油地球物理勘探,2021,56(4):688-697.YANG Wuyang,YANG Jiarun,CHEN Shuangquan,et al.Seismic data fault detection based on U-Net deep learning network[J].Oil Geophysical Prospecting,2021,56(4):688-697.

    • [17] CHEHRAZI A,RAHIMPOUR-BONAB H,REZAEE M R.Seis⁃ mic data conditioning and neural network-based attribute selec⁃ tion for enhanced fault detection[J].Petroleum Geoscience,2013,19(2):169-183.

    • [18] ZHANG C,FROGNER C,ARAYA-POLO M,et al.Machinelearning based automated fault detection in seismic traces[C].Amsterdam:76th EAGE Conference and Exbibition,2014:1-5.

    • [19] HUANG L,DONG X S,CLEE T E.A scalable deep learning plat⁃ form for identifying geologic features from seismic attributes[J].The Leading Edge,2017,36(3):249-256.

    • [20] XIONG W,JI X,MA Y,et al.Seismic fault detection with convolu⁃ tional neural network[J].Geophysics,2018,83(5):O97-O103.

    • [21] MA Yide,LIU Qing,QIAN Zhibai.Automated image segmentation using improved PCNN model based on cross-entropy[C].Hong Kong:Proceedings of 2004 International Symposium on Intelligent Multimedia,Video and Speech Processing,2004:743-746.

    • [22] LIN T Y,GOYAL P,GIRSHICK R,et al.Focal loss for dense ob⁃ ject detection[C].Venice:IEEE Transactions on Pattern Analysis and Machine Intelligence,2017:2 999-3 007.

    • [23] SUDRE C H,LI Wenqi,VERCAUTEREN Tom,et al.Generalised Dice overlap as a deep learning loss function for highly unbal⁃ anced segmentations[C].Quebec City:International Workshop on Deep Learning in Medical Image Analysis International Workshop on Multimodal Learning for Clinical Decision Support,2017:240-248.

    • [24] JADON S.A survey of loss functions for semantic segmentation [C].Viña del Mar:IEEE Conference on Computational Intelli⁃ gence in Bioinformatics and Computational Biology,2020:1-7.

    • [25] CHEN L C,PAPANDREOU G,KOKKINOS I,et al.Semantic im⁃ age segmentation with deep convolutional nets and fully connect⁃ ed CRFs[J].Computer Science,2014,(4):357-361.

    • [26] CHEN L C,PAPANDREOU G,KOKKINOS I,et al.DeepLab:Se⁃ mantic image segmentation with deep convolutional nets,atrous convolution,and fully connected CRFs[J].IEEE Transactions on Pattern Analysis and Machine Intelligence,2018,40(4):834-848.

    • [27] CHEN L C,ZHU Yukun,PAPANDREOU G,et al.Encoder-de⁃ coder with atrous separable convolution for semantic image seg⁃ mentation[C].Munich:Proceedings of the European Conference on Computer Vision,2018:801-818.

    • [28] ZHAO H S,SHI J P,QI X J,et al.Pyramid scene parsing network [C].Hawaii:IEEE Conference on Computer Vision and Pattern Recognition,2017:6 230-6 239.

    • [29] SHELHAMER E,LONG J,DARRELL T.Fully convolutional net⁃ works for semantic segmentation[J].IEEE Transactions on Pattern Analysis and Machine Intelligence,2017,39(4):640-651.

    • [30] RONNEBERGER O,FISCHER P,BROX T.U-Net:Convolutional networks for biomedical image segmentation[C].Munich:Medical Image Computing and Computer Assisted Intervention,2015:234-241.

    • [31] 秦飞巍,沈希乐,彭勇,等.无人驾驶中的场景实时语义分割方法[J].计算机辅助设计与图形学学报,2021,33(7):1 026-1 037.QIN Feiwei,SHEN Xiyue,Peng Yong,et al.A real-time semantic segmentation approach for autonomous driving scenes[J].Journal of Computer-Aided Design & Computer Graphics,2021,33(7):1 026-1 037.

    • [32] 胡嵽,冯子亮.基于深度学习的轻量级道路图像语义分割算法 [J].计算机应用,2021,41(5):1 326-1 331.HU Die,FENG Ziliang.Light-weight road image semantic seg⁃ mentation algorithm based on deep learning[J].Journal of Com⁃ puter Applications,2021,41(5):1 326-1 331.

    • [33] 方超伟,李雪,李钟毓,等.基于双模型交互学习的半监督医学图像分割[J].自动化学报,2020,46(13):1-15.FANG Chaowei,LI Xue,LI Zhongyu,et al.Interactive dual-mod⁃ el learning for semi-supervised medical image segmentation[J].Acta Automatica Sinica,2020,46(13):1-15.

    • [34] 徐世杰,杜煜,鹿鑫,等.基于ENet的轻量级语义分割算法研究 [J].计算机工程与科学,2021,43(8):1 454-1 460.XU Shijie,DU Yu,LU Xin,et al.A lightweight semantic segmen⁃ tation algorithm based on ENet[J].Computer Engineering & Sci⁃ ence,2021,43(8):1 454-1 460.

    • [35] FAN Heng,LING Haibin.SANet:Structure-aware network for vi⁃ sual tracking[C].Hawaii:IEEE Conference on Computer Vision and Pattern Recognition Workshops,2017:2 217-2 224.

    • [36] HUANG Zilong,WANG Xinggang,HUANG Lichao,et al.CCNet:Criss-cross attention for semantic segmentation[C].Seoul:IEEE/CVF International Conference on Computer Vision,2019:603-612.

    • [37] FU Jun,LIU Jing,TIAN Haijie,et al.Dual attention network for scene segmentation[C].Long Beach:IEEE/CVF Conference on Computer Vision and Pattern Recognition,2020:3 141-3 149.

    • [38] HUANG Huimin,LIN Lanfen,TONG Ruofeng,et al.UNet3+:A full-scale connected UNet for medical image segmentation[C].Barcelona:IEEE International Conference on Acoustics,Speech and Signal Processing,2020:1 055-1 059.

    • [39] ZHOU Zongwei,SIDDIQUEE M M R,TAJBAKHSH N,et al.UN⁃ et++:A nested U-Net architecture for medical image segmentation [C].Berlin:Deep Learning in Medical Image Analysis and Multi⁃ modal Learning for Clinical Decision Support,2018:3-11.

    • [40] VASWANI A,RAMACHANDRAN P,SRINIVAS A,et al.Scaling local self-attention for parameter efficient visual backbones[C].Nashville:IEEE/CVF Conference on Computer Vision and Pattern Recognition,2021:12 889-12 899.

    • [41] KINGMA D P,BA J L.Adam:A method for stochastic optimization [C].San Diego:3rd International Conference on Learning Repre⁃ sentations,2015:1-15.

  • 参考文献

    • [1] SMITH D A.Sealing and nonsealing faults in Louisiana Gulf Coast salt basin[J].AAPG Bulletin,1980,64(2):145-172.

    • [2] HOOPER E C D.Fluid migration along growth faults in compact⁃ ing sediments[J].Journal of Petroleum Geology,1991,14(S1):161-180.

    • [3] 李强,田晓平,何京,等.断层封闭性定量表征及对油气富集的控制作用[J].西南石油大学学报:自然科学版,2018,40(4):40-50.LI Qiang,TIAN Xiaoping,HE Jing,et al.Quantitative character⁃ ization of fault sealing and its control on hydrocarbon accumula⁃ tion[J].Journal of Southwest Petroleum University:Science & Technology Edition,2018,40(4):40-50.

    • [4] MARFURT K J,KIRLIN R L,FARMER S L,et al.3-D seismic at⁃ tributes using a semblance-based coherency algorithm[J].Geo⁃ physics,1998,63(4):1 150-1 165.

    • [5] 陆文凯,张善文,肖焕钦.基于相干滤波的相干体图像增强[J].天然气工业,2006,26(5):37-39.LU Wenkai,ZHANG Shanwen,XIAO Huanqin.Image enhance⁃ ment of coherence cube based on coherent filtering[J].Natural Gas Industry,2006,26(5):37-39.

    • [6] 马承杰.多尺度边缘检测技术在断层识别及裂缝发育带预测中的应用——以车排子地区排691井区为例[J].油气地质与采收率,2021,28(2):85-90.MA Chengjie.Application of multi-scale edge detection technolo⁃ gy to fault recognition and fracture zone prediction:A case study of Block Well P691,Chepaizi area[J].Petroleum Geology and Re⁃ covery Efficiency,2021,28(2):85-90.

    • [7] RANDEN T,PEDERSEN S I,SØNNELAND L.Automatic extrac⁃ tion of fault surfaces from three-dimensional seismic data[C].San Antonio:SEG International Exposition and Annual Meeting,2001:551-554.

    • [8] ROBERTS A.Curvature attributes and their application to 3D in⁃ terpreted horizons[J].First Break,2001,19(2):85-100.

    • [9] DI H B,GAO D.Efficient volumetric extraction of most positive/negative curvature and flexure for fracture characterization from 3D seismic data[J].Geophysical Prospecting,2016,64(6):1 454-1 468.

    • [10] HALE D.Methods to compute fault images,extract fault surfaces,and estimate fault throws from 3D seismic images[J].Geophysics,2013,78(2):O33-O43.

    • [11] PEDERSEN S I,RANDEN T,SONNELAND L,et al.Automatic fault extraction using artificial ants[C].Salt Lake City:72nd An⁃ nual International Meeting,SEG,Expanded Abstracts,2002.

    • [12] PEDERSEN S I,SKOV T,HETLELID A,et al.New paradigm of fault interpretation[C].Dallas:73rd Annual International Meet⁃ ing,SEG,Expanded Abstracts,2003.

    • [13] WU Xinming,ZHU Zhihui.Methods to enhance seismic faults and construct fault surfaces[J].Computers & Geosciences,2017,107:37-48.

    • [14] WU Xinming,FOMEL S.Automatic fault interpretation with opti⁃ mal surface voting[J].Geophysics,2018,83(5):O67-O82.

    • [15] WU Xinming,LIANG Luming,SHI Yunzhi,et al.FaultSeg3D:Us⁃ ing synthetic datasets to train an end-to-end convolutional neural network for 3D seismic fault segmentation[J].Geophysics,2019,84(3):35-45.

    • [16] 杨午阳,杨佳润,陈双全,等.基于U-Net深度学习网络的地震数据断层检测[J].石油地球物理勘探,2021,56(4):688-697.YANG Wuyang,YANG Jiarun,CHEN Shuangquan,et al.Seismic data fault detection based on U-Net deep learning network[J].Oil Geophysical Prospecting,2021,56(4):688-697.

    • [17] CHEHRAZI A,RAHIMPOUR-BONAB H,REZAEE M R.Seis⁃ mic data conditioning and neural network-based attribute selec⁃ tion for enhanced fault detection[J].Petroleum Geoscience,2013,19(2):169-183.

    • [18] ZHANG C,FROGNER C,ARAYA-POLO M,et al.Machinelearning based automated fault detection in seismic traces[C].Amsterdam:76th EAGE Conference and Exbibition,2014:1-5.

    • [19] HUANG L,DONG X S,CLEE T E.A scalable deep learning plat⁃ form for identifying geologic features from seismic attributes[J].The Leading Edge,2017,36(3):249-256.

    • [20] XIONG W,JI X,MA Y,et al.Seismic fault detection with convolu⁃ tional neural network[J].Geophysics,2018,83(5):O97-O103.

    • [21] MA Yide,LIU Qing,QIAN Zhibai.Automated image segmentation using improved PCNN model based on cross-entropy[C].Hong Kong:Proceedings of 2004 International Symposium on Intelligent Multimedia,Video and Speech Processing,2004:743-746.

    • [22] LIN T Y,GOYAL P,GIRSHICK R,et al.Focal loss for dense ob⁃ ject detection[C].Venice:IEEE Transactions on Pattern Analysis and Machine Intelligence,2017:2 999-3 007.

    • [23] SUDRE C H,LI Wenqi,VERCAUTEREN Tom,et al.Generalised Dice overlap as a deep learning loss function for highly unbal⁃ anced segmentations[C].Quebec City:International Workshop on Deep Learning in Medical Image Analysis International Workshop on Multimodal Learning for Clinical Decision Support,2017:240-248.

    • [24] JADON S.A survey of loss functions for semantic segmentation [C].Viña del Mar:IEEE Conference on Computational Intelli⁃ gence in Bioinformatics and Computational Biology,2020:1-7.

    • [25] CHEN L C,PAPANDREOU G,KOKKINOS I,et al.Semantic im⁃ age segmentation with deep convolutional nets and fully connect⁃ ed CRFs[J].Computer Science,2014,(4):357-361.

    • [26] CHEN L C,PAPANDREOU G,KOKKINOS I,et al.DeepLab:Se⁃ mantic image segmentation with deep convolutional nets,atrous convolution,and fully connected CRFs[J].IEEE Transactions on Pattern Analysis and Machine Intelligence,2018,40(4):834-848.

    • [27] CHEN L C,ZHU Yukun,PAPANDREOU G,et al.Encoder-de⁃ coder with atrous separable convolution for semantic image seg⁃ mentation[C].Munich:Proceedings of the European Conference on Computer Vision,2018:801-818.

    • [28] ZHAO H S,SHI J P,QI X J,et al.Pyramid scene parsing network [C].Hawaii:IEEE Conference on Computer Vision and Pattern Recognition,2017:6 230-6 239.

    • [29] SHELHAMER E,LONG J,DARRELL T.Fully convolutional net⁃ works for semantic segmentation[J].IEEE Transactions on Pattern Analysis and Machine Intelligence,2017,39(4):640-651.

    • [30] RONNEBERGER O,FISCHER P,BROX T.U-Net:Convolutional networks for biomedical image segmentation[C].Munich:Medical Image Computing and Computer Assisted Intervention,2015:234-241.

    • [31] 秦飞巍,沈希乐,彭勇,等.无人驾驶中的场景实时语义分割方法[J].计算机辅助设计与图形学学报,2021,33(7):1 026-1 037.QIN Feiwei,SHEN Xiyue,Peng Yong,et al.A real-time semantic segmentation approach for autonomous driving scenes[J].Journal of Computer-Aided Design & Computer Graphics,2021,33(7):1 026-1 037.

    • [32] 胡嵽,冯子亮.基于深度学习的轻量级道路图像语义分割算法 [J].计算机应用,2021,41(5):1 326-1 331.HU Die,FENG Ziliang.Light-weight road image semantic seg⁃ mentation algorithm based on deep learning[J].Journal of Com⁃ puter Applications,2021,41(5):1 326-1 331.

    • [33] 方超伟,李雪,李钟毓,等.基于双模型交互学习的半监督医学图像分割[J].自动化学报,2020,46(13):1-15.FANG Chaowei,LI Xue,LI Zhongyu,et al.Interactive dual-mod⁃ el learning for semi-supervised medical image segmentation[J].Acta Automatica Sinica,2020,46(13):1-15.

    • [34] 徐世杰,杜煜,鹿鑫,等.基于ENet的轻量级语义分割算法研究 [J].计算机工程与科学,2021,43(8):1 454-1 460.XU Shijie,DU Yu,LU Xin,et al.A lightweight semantic segmen⁃ tation algorithm based on ENet[J].Computer Engineering & Sci⁃ ence,2021,43(8):1 454-1 460.

    • [35] FAN Heng,LING Haibin.SANet:Structure-aware network for vi⁃ sual tracking[C].Hawaii:IEEE Conference on Computer Vision and Pattern Recognition Workshops,2017:2 217-2 224.

    • [36] HUANG Zilong,WANG Xinggang,HUANG Lichao,et al.CCNet:Criss-cross attention for semantic segmentation[C].Seoul:IEEE/CVF International Conference on Computer Vision,2019:603-612.

    • [37] FU Jun,LIU Jing,TIAN Haijie,et al.Dual attention network for scene segmentation[C].Long Beach:IEEE/CVF Conference on Computer Vision and Pattern Recognition,2020:3 141-3 149.

    • [38] HUANG Huimin,LIN Lanfen,TONG Ruofeng,et al.UNet3+:A full-scale connected UNet for medical image segmentation[C].Barcelona:IEEE International Conference on Acoustics,Speech and Signal Processing,2020:1 055-1 059.

    • [39] ZHOU Zongwei,SIDDIQUEE M M R,TAJBAKHSH N,et al.UN⁃ et++:A nested U-Net architecture for medical image segmentation [C].Berlin:Deep Learning in Medical Image Analysis and Multi⁃ modal Learning for Clinical Decision Support,2018:3-11.

    • [40] VASWANI A,RAMACHANDRAN P,SRINIVAS A,et al.Scaling local self-attention for parameter efficient visual backbones[C].Nashville:IEEE/CVF Conference on Computer Vision and Pattern Recognition,2021:12 889-12 899.

    • [41] KINGMA D P,BA J L.Adam:A method for stochastic optimization [C].San Diego:3rd International Conference on Learning Repre⁃ sentations,2015:1-15.