改进Unet网络对叠后地震数据的断层识别

doi:10.19596/j.cnki.1001-246x.8684

摘要/Abstract

摘要：

为了提高断层识别的准确率, 提出改进Unet模型。为编码器部分设计一种多分支的并联结构M-block(Multi-branch block), 它可以捕获多尺度上下文信息, 并且多分支的并联结构会带来高性能收益。在解码器部分加入Self-Attention块和注意力门控机制。Self-Attention通过对输入特征上下文的加权平均操作, 不仅使注意力模块能够灵活地关注图像的不同区域, 而且弥补了CNN(Convolutional Neural Network)网络局部性的缺点, 为神经网络带来更多的可能性。通过合成数据和实际数据证实, 该模型将传统卷积中的权值共享优点和Self-Attention动态计算注意力权重的优点结合, 提高了断层识别的精度, 与Unet相比, 验证损失下降了33.68%。模型不仅准确识别出了断层特征, 且比目前流行的深度学习方法更准确。

关键词: 断层识别, 多分枝, Self-Attention, Unet, 编码器

Abstract:

In order to improve the accuracy of fault identification, an improved Unet model is proposed. A multi-branch parallel structure M-block (Multi-branch block) is designed for the encoder part. It can capture multi-scale context information, and multi branch parallel structure will bring high performance benefits. Self-Attention block and attention gating mechanism are added to the decoder. Self-Attention not only enables the attention module to flexibly focus on different areas of the image, but also makes up for the shortcomings of the local CNN (Convolutional Neural Network) and brings more possibilities to the neural network through the weighted average operation of the input feature context. It is verified by synthetic data and actual data that the model combines the advantages of weight sharing in traditional convolution with the advantages of Self Attention's dynamic calculation of Attention weight to improve the accuracy of fault identification. Compared with Unet, the test loss is reduced by 33.68%. The model not only identifies fault features accurately, but also is more accurate than the current popular depth learning method.

Key words: fault identification, many branches, self-attention, Unet, encoder

中图分类号:

O469

刘贵鑫, 马中华. 改进Unet网络对叠后地震数据的断层识别[J]. 计算物理, 2023, 40(6): 742-751.

Guixin LIU, Zhonghua MA. Fault Identification of Post Stack Seismic Data by Improved Unet Network[J]. Chinese Journal of Computational Physics, 2023, 40(6): 742-751.

图/表 14

图1 多分支的Self-Attention Eunet结构

Fig.1 Multibranched Self-Attention Eunet structure

图2 多分支结构

Fig.2 Multiple-branching structure

图3 Unet与M-Unet的(a)训练损失和(b)验证损失

Fig.3 (a) Training losses and (b) validation losses of Unet and M-Unet

图4 改进的Self-Attention结构

Fig.4 An improved Self-Attention structure

图5 从测试集中选择2张图片，输入网络

Fig.5 Select 2 pictures in the test set and input them to the network

图6 左侧为Unet网络的特征图，蓝色虚线框内的为加入Self-Attention后M-SA-Unet的特征图

Fig.6 The feature diagram of the Unet network on the left, and the feature diagram of M-SA-Unet after adding Self-Attention in the blue dotted box

图7 (a) 从Unet网络选择的5个epoch特征图; (b)改进M-SA-Unet网络的特征图

Fig.7 (a) Five epoch feature diagrams selected from the Unet network, and (b) the feature diagram of the improved M-SA-Unet network

图8 图像增强

Fig.8 Image enhancement

图9 在测试集上M-SA-Unet和M-SA-Eunet与其他模型断层识别的比较

Fig.9 Comparison of M-SA-Unet and M-SA-Eunet with other model fault recognition results in the test set

图10 在预测帧上M-SA-Unet和M-SA-Eunet与其他模型断层识别的比较

Fig.10 Comparison of M-SA-Unet and M-SA-Eunet with other model fault recognition results on the forecast frame

表1 在训练和验证过程中各项参数的设置和结果

Table 1 Sets and results of parameters in the training and verification process

	Train Loss	Val Loss	Epoch	Batchsize	优化器	学习率	训练miou	测试miou
Unet	0.078	0.095	30	16	AdamW	0.001	0.831	0.800
Eunet	0.058	0.091	30	16	AdamW	0.001	0.873	0.801
Segnet	0.055	0.069	30	16	AdamW	0.001	0.890	0.875
Unet++	0.035	0.065	30	16	AdamW	0.001	0.911	0.904
Unet+++	0.065	0.113	30	8	AdamW	0.001	0.848	0.749
Attention-Unet	0.055	0.076	30	16	AdamW	0.001	0.911	0.877
M-SA-Unet	0.033	0.062	30	16	AdamW	0.001	0.914	0.882
M-SA-Eunet	0.034	0.063	30	16	AdamW	0.001	0.926	0.911

图11 Unet，Eunet，M-SA-Unet和M-SA-Eunet的(a)训练损失、(b)验证损失、(c)训练miou和(d)验证miou

Fig.11 Unet, Eunet, M-SA-Unet and M-SA-Eunet (a) training loss, (b) validation loss, (c) training miou and (d) validation miou

图12 (a) 断层标签数据与(b)实际地震数据

Fig.12 (a) Fault label data and (b) actual seismic data

图13 M-SA-Eunet与其他模型在实际地震预测数据集上的比较

Fig.13 Comparison of M-SA-Eunet and other models on actual earthquake prediction data sets

参考文献 21

1	YUAN Y . Some new progress in the fields of computational petroleum geology and others[J]. Chinese Journal of Computational Physics, 2003, (4): 283- 290. DOI
2	RONNEBERGER O, FISCHER P, BROX T. U-net: Convolutional networks for biomedical image segmentation[M]//Lecture Notes in Computer Science. Cham: Springer International Publishing, 2015: 234-241.
3	ZHANG H, GOODFELLOW I, METAXAS D, et al. Self-attention generative adversarial networks[C]//International Conference on Machine Learning. PMLR, 2019: 7354-7363.
4	何易龙, 文晓涛, 王锦涛, 等. 基于3D U-Net++L~3卷积神经网络的断层识别[J]. 地球物理学进展, 2022, 37 (2): 607- 616.
5	芦凤明, 孟瑞刚, 张军华, 等. UNet++和迁移学习相结合的复杂断裂识别方法研究[J]. 地球物理学进展, 2022, 37 (3): 1100- 1111.
6	ZHOU Z W, RAHMAN S M M, TAJBAKHSH N, et al. UNet++: A nested U-net architecture for medical image segmentation[M]//Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support. Cham: Springer International Publishing, 2018: 3-11.
7	席英杰, 李克文, 徐延辉, 等. 一种用于地震断层图像识别的SPD-UNet模型[J]. 计算机工程, 2021, 47 (12): 249- 255.
8	杨晶, 丁仁伟, 林年添, 等. 基于深度学习的地震断层智能识别研究进展[J]. 地球物理学进展, 2022, 37 (1): 298- 311.
9	WANG J , ZHANG J H , ZHANG J L , et al. Research on fault recognition method combining 3D Res-UNet and knowledge distillation[J]. Applied Geophysics, 2021, 18 (2): 199- 212. DOI
10	周东红, 李辉, 阎建国. CNN标签数据迭代优化及在小断层识别中的应用[J]. 地球物理学进展, 2022, 37 (1): 338- 347.
11	WU X , LIANG L , SHI Y , et al. FaultSeg3D: Using synthetic data sets to train an end-to-end convolutional neural network for 3D seismic fault segmentation[J]. Geophysics, 2019, 84 (3): IM35- IM45. DOI
12	SZEGEDY C, VANHOUCKE V, IOFFE S, et al. Rethinking the inception architecture for computer vision[C]//2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016: 2818-2826.
13	XU X J , LIU C , ZHENG Y Y . 3D tooth segmentation and labeling using deep convolutional neural networks[J]. IEEE Transactions on Visualization and Computer Graphics, 2019, 25 (7): 2336- 2348. DOI
14	LU X , HUANG H , LI S , et al. Salt-body classification method based on U-Net[J]. Chinese Journal of Computational Physics, 2020, 37 (3): 327- 334.
15	娄莉, 张丰侠, 韩柏迅. 基于图像分割的盐丘识别算法研究[J]. 计算物理, 2023, 40 (4): 511- 518.
16	ÇIÇEK Ö, ABDULKADIR A, LIENKAMP S S, et al. 3D U-net: Learning dense volumetric segmentation from sparse annotation[M]//Medical Image Computing and Computer-assisted Intervention-MICCAI 2016. Cham: Springer International Publishing, 2016: 424-432.
17	LI X M , CHEN H , QI X J , et al. H-DenseUNet: Hybrid densely connected UNet for liver and tumor segmentation from CT volumes[J]. IEEE Transactions on Medical Imaging, 2018, 37 (12): 2663- 2674.
18	BADRINARAYANAN V , KENDALL A , CIPOLLA R . SegNet: A deep convolutional encoder-decoder architecture for image segmentation[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017, 39 (12): 2481- 2495.
19	HUANG H M, LIN L F, TONG R F, et al. UNet 3: A full-scale connected UNet for medical image segmentation[C]//ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2020: 1055-1059.
20	SZEGEDY C, LIU W, JIA Y Q, et al. Going deeper with convolutions[C]//2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015: 1-9.
21	ASHISH V , NOAM S , NIKI P , et al. Attention is all you need[J]. Advances in Neural Information Processing Systems, 2017, 30, 5998- 6008.

[1]	郭城, 程鹏丹, 王亚辉. 修正模板改进的五阶WENO-Z+格式[J]. 计算物理, 2023, 40(6): 666-676.
[2]	李鹏迪, 刘俊, 郑淇蓉, 张传国, 李永钢, 张永胜, 赵高峰, 曾雉. 硅中硼离子注入的带电缺陷动力学模拟[J]. 计算物理, 2021, 38(3): 361-370.
[3]	刘涛, 杨子义, 陈雨青, 高涛. 重费米子超导PuMGa₅(M=Co,Rh)结构、电子和热力学性质的第一性原理研究[J]. 计算物理, 2021, 38(1): 106-112.
[4]	王书松, 张素英. 谐振子势与高斯势联合势阱中玻色爱因斯坦凝聚体的巨涡旋态[J]. 计算物理, 2021, 38(1): 113-119.
[5]	刘永智, 解立强, 梁盛德, 朱开礼, 席忠红, 袁方强. C₆₀对细胞膜潜在生物毒性的分子动力学模拟[J]. 计算物理, 2020, 37(4): 479-487.
[6]	魏志超, 王能平. 介质中双缺陷电荷对碳纳米管场效应晶体管量子输运特性的影响[J]. 计算物理, 2020, 37(3): 352-364.
[7]	张雪颖, 冯琳. C掺杂Mn₃Ge的电子结构和磁性[J]. 计算物理, 2019, 36(6): 742-748.
[8]	任娟, 张宁超, 刘萍萍. 碳气凝胶储氢性能的理论研究[J]. 计算物理, 2019, 36(6): 749-756.
[9]	唐攀飞, 郑淇蓉, 李敬文, 魏留明, 张传国, 李永钢, 曾雉. 计及级联内缺陷空间关联的团簇动力学模拟[J]. 计算物理, 2019, 36(5): 586-594.
[10]	尹海峰, 曾春花, 陈文经. 二维二元碳化硅纳米结构的等离激元激发[J]. 计算物理, 2019, 36(5): 603-609.
[11]	马文杰, 张平. 计算顺磁-铁磁相变临界磁场的一种方法[J]. 计算物理, 2019, 36(4): 421-426.
[12]	莫康信, 苏佳佳. 磁性纳米颗粒系统偶极相互作用的Monte Carlo研究[J]. 计算物理, 2019, 36(3): 335-341.
[13]	张海燕, 殷新春. 简单金属固液界面固化过程生长机制的分子动力学研究[J]. 计算物理, 2019, 36(1): 80-88.
[14]	马瑞, 张华林. 掺杂菱形BN片的石墨烯纳米带的电子特性[J]. 计算物理, 2019, 36(1): 99-105.
[15]	周康, 冯庆, 田芸, 李科, 周清斌. 过渡金属Cu、Cr掺杂TiO₂表面氧化性气体NO₂光学气敏传感特性[J]. 计算物理, 2018, 35(6): 702-710.