基于3D-CA-GAN的岩石体纹理合成技术

段炼; 冯云; 花卫华; 陈启浩; 刘修国; 张坤; 付伟

doi:10.3799/dqkx.2025.134

基于3D-CA-GAN的岩石体纹理合成技术

doi: 10.3799/dqkx.2025.134

段炼^1, ,,
冯云¹,
花卫华^1, , ,,
陈启浩¹,
刘修国¹,
张坤²,
付伟²

1.
中国地质大学(武汉)地理与信息工程学院, 湖北武汉 430078
2.
中铁第一勘察设计院集团有限公司, 陕西西安 710043

基金项目:

中铁第一勘察设计院集团有限公司科研项目 2022KY53ZD(CYH)-10

中国铁建股份有限公司重大专项 2024-W04

详细信息

作者简介:
段炼（2002-），女，研究生，主要研究领域为三维模型可视化. ORCID：0009-0009-8628-5222. E-mail：duanlian@cug.edu.cn

通讯作者:
花卫华，ORCID：0000-0002-2255-7411. E-mail: huaweihua@cug.edu.cn

中图分类号: P58
计量
- 文章访问数: 107
- HTML全文浏览量: 19
- PDF下载量: 5
- 被引次数: 0
出版历程
- 收稿日期: 2025-04-15
- 刊出日期: 2025-11-25

Rock Solid Texture Synthesis Based on 3D-CA-GAN

Duan Lian^{1
,
,},
Feng Yun¹,
Hua Weihua^{1
, ,
,},
Chen Qihao¹,
Liu Xiuguo¹,
Zhang Kun²,
Fu Wei²

1.
School of Geography and Information Engineering, China University of Geosciences (Wuhan), Wuhan 430078, China
2.
China Railway First Survey and Design Institute Group Co. Ltd., Xi'an 710043

摘要

摘要: 基于二维样本（深度学习）的体纹理合成是一种重要的岩石体纹理生成途径，目前岩石体纹理合成存在无法长距离依赖和颜色失真的问题.提出一种基于三维坐标注意力生成对抗网络（3D-Coordinate Attention Generative Adversarial Network，简称3D-CA-GAN）的创新方法.通过将坐标注意力机制（Coordinate Attention，简称CA）扩展至三维空间，结合内容感知上采样模块和多尺度判别器，实现了矿物颗粒空间分布的高保真建模.实验表明，该方法在SSIM（0.773）、PSNR（提升24.92%）和LPIPS（降低0.110）等指标上显著优于现有技术，消融实验进一步验证3D-CA模块使方向性纹理的SSIM提升14.69%.本研究为地质建模提供了具有真实感纹理合成的新解决方案，其三维注意力框架对通用生成任务具有借鉴意义.
- 岩石 /
- 体纹理 /
- 混合空洞卷积 /
- 注意力模块 /
- 3D-CA-GAN /
- 三维建模
Abstract: Solid texture synthesis based on 2D samples (deep learning) is an important pathway for rock solid texture generation, which currently suffers from the inability of long distance dependence and color distortion. In this paper, it proposes an innovative method based on 3D coordinate attention generative adversarial network (3D-CA-GAN). By extending the coordinate attention mechanism to three-dimensional space (3D-CA) and combining the content-aware upsampling module and multi-scale discriminator, high-fidelity modeling of the spatial distribution of mineral particles is achieved. Experiments show that the method significantly outperforms existing techniques in terms of SSIM (0.773), PSNR (24.92% enhancement), and LPIPS (0.110 reduction), and ablation experiments further validate that the 3D-CA module improves the SSIM of directional textures by 14.69%. This study provides a new solution to texture synthesis with realism for geological modeling, and its 3D attention framework is useful for generic generation tasks.
- rocks /
- solid texture /
- hybrid dilated convolution /
- attention module /
- 3D-CA-GAN /
- 3d-modeling

HTML全文

图 1 3D-CA-GAN总体框架

Fig. 1. Overall frame diagram of 3D-CA-GAN

下载: 全尺寸图片幻灯片

图 2 生成器网络架构

Fig. 2. Generator network architecture

下载: 全尺寸图片幻灯片

图 3 多尺度鉴别器结构

Fig. 3. Multi-scale discriminator structure

下载: 全尺寸图片幻灯片

图 4 混合空洞卷积的扩张感受野与网格效应分析示意

Fig. 4. Diagram of the dilated receptive field and grid effect analysis in hybrid dilated convolution

下载: 全尺寸图片幻灯片

图 5 混合空洞卷积块

Fig. 5. Hybrid dilated convolution block

下载: 全尺寸图片幻灯片

图 6 3D-Coordinate attention模块

Fig. 6. 3D coordinate attention module

下载: 全尺寸图片幻灯片

图 7 内容感知的上采样模块

Fig. 7. Content-aware upsampling module

下载: 全尺寸图片幻灯片

图 8 部分数据集构建流程

Fig. 8. Partial dataset construction pipeline

下载: 全尺寸图片幻灯片

图 9 岩石纹理图片

Fig. 9. Rock texture images

下载: 全尺寸图片幻灯片

图 10 岩石体纹理生成结果展示

Fig. 10. Synthetic solid volume texture exemplars

下载: 全尺寸图片幻灯片

图 11 岩石体纹理的训练损失值展示

Fig. 11. Loss convergence profile of solid volume texture synthesis

下载: 全尺寸图片幻灯片

图 12 岩石体纹理在3D网格模型上的纹理映射效果

Fig. 12. 3D mesh texture mapping visualization of synthesized solid volumes

下载: 全尺寸图片幻灯片

图 13 岩石体纹理在经典3D网格模型上的纹理映射效果

Fig. 13. UV-mapping of synthesized solid textures on standard 3D mesh primitives

下载: 全尺寸图片幻灯片

图 14 非均匀岩石体纹理合成效果

Fig. 14. Non-uniform solid volume texture synthesis results

下载: 全尺寸图片幻灯片

图 15 定性对比实验结果

Fig. 15. Qualitative comparison results

下载: 全尺寸图片幻灯片

图 16 不同迭代次数下的对比实验结果

Fig. 16. Comparative experimental results under different iteration counts

下载: 全尺寸图片幻灯片

图 17 各向异性体纹理的对比实验

Fig. 17. Anisotropic solid texture simulation results

下载: 全尺寸图片幻灯片

图 18 消融实验结果

Fig. 18. Ablation study results

下载: 全尺寸图片幻灯片

表 1 硬件设施

Table 1. Hardware facilities

名称	规格参数
显卡（GPU）	NVIDIA Quadro RTX 6000
处理器（CPU）	Intel(R) Core(TM) i9-10900K CPU @3.70GHz 3.70
显存	24G

下载: 导出CSV

表 2 软件配置

Table 2. Software configurations

名称	版本
操作系统	Windows 10
深度学习框架	pytorch
CUDA/cuDNN	Cuda11.3
Python	3.8

下载: 导出CSV

表 3 模型训练参数

Table 3. Model training parameters

训练参数	描述	参数值
scale_k	学习的尺度	5
batch_size	批次大小	1
learning_g	生成器学习率	5e-4
learning_d	鉴别器学习率	3e-4

下载: 导出CSV

表 4 定量对比实验结果

Table 4. Quantitative comparison results

样本	方法	SSIM ↑	PNSR ↑	LPIPS ↓	FID ↓	DISTS ↓
	CNN	0.460	13.682	0.373	27.3	0.242
a	STS-GAN	0.701	13.467	0.274	22.8	0.253
	ours	0.773	14.678	0.263	11.0	0.156
	CNN	0.395	12.874	0.463	26.3	0.278
b	STS-GAN	0.689	14.287	0.398	23.6	0.257
	ours	0.621	15.014	0.302	12.3	0.203
注：↑代表更高更好；↓代表更低更好.

下载: 导出CSV

表 5 定量消融实验结果

Table 5. Quantitative ablation study results

样本	方法	SSIM ↑	PNSR ↑	LPIPS ↓	DISTS ↓
	STS-GAN	0.701	13.467	0.274	0.253
(a)	+CA注意力	0.804	16.823	0.279	0.269
	+上采样模块	0.722	15.327	0.250	0.166
	+CA注意力+上采样模块	0.773	14.678	0.263	0.156
	STS-GAN	0.689	14.287	0.398	0.257
(b)	+CA注意力	0.726	15.288	0.316	0.236
	+上采样模块	0.616	14.936	0.437	0.313
	+CA注意力+上采样模块	0.621	15.014	0.302	0.203
注：↑代表更高更好；↓代表更低更好.

下载: 导出CSV

表 6 混合空洞卷积块消融实验结果

Table 6. Ablation study results of hybrid dilated convolution blocks

样本	方法	参数量	训练时长（h）
(a)	普通卷积块	68 227	8.5
(a)	混合空洞卷积块	44 723	6.2

下载: 导出CSV

参考文献(29)

Cao, T. S., Kreis, K., Fidler, S., et al., 2023. TexFusion: Synthesizing 3D Textures with Text-Guided Image Diffusion Models. IEEE International Conference on Computer Vision (ICCV), Paris, 4169-4181. https://arxiv.org/abs/2310.13772
Du, S. P., Hu, S. M., Martin, R. R., 2013. Semiregular Solid Texturing from 2D Image Exemplars. IEEE Transactions on Visualization and Computer Graphics, 19(3): 460-469. https://doi.org/10.1109/TVCG.2012.129
Fu, J. M., Hu, M. S., Fang, F., et al., 2024. Complex Orebody 3D Modeling Using Radial Basis Function Surface Incorporating Stacking Integration Strategy. Earth Science, 49(3): 1165-1176 (in Chinese with English abstract).
Gatys, L. A., Ecker, A. S., Bethge, M., 2015. Texture Synthesis Using Convolutional Neural Networks. Advances in Neural Information Processing Systems (NIPS), La Jolla, 1-9. https://doi.org/10.5555/2969239.2969269
Gatys, L. A., Ecker, A. S., Bethge, M., 2016. Image Style Transfer Using Convolutional Neural Networks. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, 2414-2423. https://doi.org/10.1109/CVPR.2016.265
Gulrajani, I., Ahmed, F., Arjovcky, M., et al., 2017. Improved Training of Wasserstein GANs. Proceedings of the 31st International Conference on Neural Information Processing Systems, Long Beach, 5769-5779. https://doi.org/10.5555/3295222.3295327
Gutierrez, J., Rabin, J., Galerne, B., et al., 2020. On Demand Solid Texture Synthesis Using Deep 3D Networks. Computer Graphics Forum, 39(1): 511-530. https://doi.org/10.1111/cgf.13889
Heeger, D. J., Bergen, J. R., 1995. Pyramid-Based Texture Analysis/Synthesis. Proceedings of the 22nd Annual Conference on Computer Graphics and Interactive Techniques, New York, 229-238. https://doi.org/10.1145/218380.218446
Henzler, P., Mitra, N. J., Ritschel, T., 2020. Learning a Neural 3D Texture Space from 2D Exemplars. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Seattle, 8353-8361. https://doi.org/10.1109/cvpr42600.2020.00838
Hou, Q. B., Zhou, D. Q., Feng, J. S., 2021. Coordinate Attention for Efficient Mobile Network Design. Proceedings of the 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Nashville, 13708-13717. https://doi.org/10.1109/CVPR46437.2021.01350
Hua, W. H., Xiao, Y. N., Wang, Z. J., et al., 2022. Real-Time Generation Technology of Vector Geological Profile Based on 3D Geological Model. Earth Science, 47(11): 4256-4266 (in Chinese with English abstract).
Huo, D., Guo, Z. X., Zuo, X. X., et al., 2024. TexGen: Text-Guided 3D Texture Generation with Multi-View Sampling and Resampling. In: Leonardis, A., Ricci, E., Roth, S., eds., Computer Vision-ECCV 2024. Springer, Cham, 352-368. https://doi.org/10.1007/978-3-031-72920-1_20
Kwatra, V., Essa, I., Bobick, A., et al., 2005. Texture Optimization for Example-Based Synthesis. ACM SIGGRAPH 2005 Papers, Los Angeles, 795-802. https://doi.org/10.1145/1186822.1073263
Peachey, D. R., 1985. Solid Texturing of Complex Surfaces. ACM SIGGRAPH Computer Graphics, 19(3): 279-286. https://doi.org/10.1145/325165.325246
Perlin, K., 1985. An Image Synthesizer. ACM SIGGRAPH Computer Graphics, 19(3): 287-296. https://doi.org/10.1145/325165.325247
Portenier, T., Bigdeli, S., Goksel, O., 2020. GramGAN: Deep 3D Texture Synthesis from 2D Exemplars. Proceedings of the 34th International Conference on Neural Information Processing Systems, Montreal, 6994-7004. https://doi.org/10.5555/3495724.3496311
Qian, Y. L., Shi, J., Sun, H. Q., et al., 2023. Vector Solid Texture Synthesis Using Unified RBF-Based Representation and Optimization. The Visual Computer, 39(9): 3963-3977. https://doi.org/10.1007/s00371-022-02541-y
Qian, Y. L., Shu, Y., Sun, H. Q., et al., 2015. Vector Solid Texture Synthesis Using Two-Scale Shaping Model. Proceedings of the 21st ACM Symposium on Virtual Reality Software and Technology, Beijing, 27-36. https://doi.org/10.1145/2821592.2821605
Shaham, T. R., Dekel, T., Michaeli, T., 2019. SinGAN: Learning a Generative Model from a Single Natural Image. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), Seoul, 4569-4579. https://doi.org/10.1109/iccv.2019.00467
Tai, W. X., Zhou, Q., Yang, C. F., et al., 2023.3D Geological Visualization Modeling and Its Application in Zhexiang Gold Deposit, Southwest Guizhou Province. Earth Science, 48(11): 4017-4033 (in Chinese with English abstract).
Wang, J. Q., Chen, K., Xu, R., et al., 2022. CARAFE: Unified Content-Aware Re-Assembly of FEatures. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(9): 4674-4687. https://doi.org/10.1109/TPAMI.2021.3074370
Wang, P. Q., Chen, P. F., Yuan, Y., et al., 2018. Understanding Convolution for Semantic Segmentation. 2018 IEEE Winter Conference on Applications of Computer Vision (WACV), Lake Tahoe, 1451-1460. https://doi.org/10.1109/WACV.2018.00163
Xiao, H. G., He, L., Zheng, Y. L., et al., 2022.3D Solid Digital and Numerical Modeling of Multimineral Heterogeneous Rocks Based on Deep Learning. Geomechanics and Geophysics for Geo-Energy and Geo-Resources, 8(6): 188. https://doi.org/10.1007/s40948-022-00495-y
Zirek, S., 2023. Synthesising 3D Solid Models of Natural Heterogeneous Materials from Single Sample Image, Using Encoding Deep Convolutional Generative Adversarial Networks. Systems and Soft Computing, 5: 200051. https://doi.org/10.1016/j.sasc.2023.200051
Zhang, R., Isola, P., Efros, A. A., et al., 2018. The Unreasonable Effectiveness of Deep Features as a Perceptual Metric. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, Salt Lake City, 586-595. https://doi.org/10.1109/CVPR.2018.00068
Zhao, X., Guo, J. F., Wang, L., et al., 2023. STS-GAN: Can We Synthesize Solid Texture with High Fidelity from Arbitrary 2D Exemplar? In: Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence (IJCAI). International Joint Conferences on Artificial Intelligence Organization, Macao, 1768-1776. https://doi.org/10.24963/ijcai.2023/196
扶金铭, 胡茂胜, 方芳, 等, 2024. Stacking集成策略下的径向基函数曲面复杂矿体三维建模方法. 地球科学, 49(3): 1165-1176. doi: 10.3799/dqkx.2022.433
花卫华, 肖旖旎, 王振娟, 等, 2022. 基于三维地质模型的矢量地质剖面图实时生成技术. 地球科学, 47(11): 4256-4266. doi: 10.3799/dqkx.2022.291
邰文星, 周琦, 杨成富, 等, 2023. 黔西南者相金矿床三维地质可视化建模及应用. 地球科学, 48(11): 4017-4033. doi: 10.3799/dqkx.2022.095