结合神经网络和奇异值分解的单图像材质重建

doi:10.16182/j.issn1004731x.joss.25-0834

系统仿真学报 ›› 2026, Vol. 38 ›› Issue (1): 189-199.doi: 10.16182/j.issn1004731x.joss.25-0834

结合神经网络和奇异值分解的单图像材质重建

李志强¹, 沈旭昆², 胡勇², 周雪杨², 陈弈帆¹

^1.哈尔滨工程大学电子政务建模仿真国家工程实验室，黑龙江哈尔滨 150001
^2.北京航空航天大学虚拟现实技术与系统全国重点实验室，北京 100191

收稿日期:2025-09-02 修回日期:2025-12-03 出版日期:2026-01-18 发布日期:2026-01-28
通讯作者: 胡勇
第一作者简介:李志强(1991-)，男，白族，讲师，博士，研究方向为虚拟现实。
基金资助:
国家自然科学基金(62572143);虚拟现实技术与系统全国重点实验室(北京航空航天大学)开放课题基金(VRLAB2025C18)

Material Reconstruction from Single Image Combining Neural Networks with Singular Value Decomposition

Li Zhiqiang¹, Shen Xukun², Hu Yong², Zhou Xueyang², Chen Yifan¹

^1.National Engineering Laboratory for Modeling and Emulation in E-Government, Harbin Engineering University, Harbin 150001, China
^2.State Key Laboratory of Virtual Reality Technology and Systems, Beihang University, Beijing 100191, China

Received:2025-09-02 Revised:2025-12-03 Online:2026-01-18 Published:2026-01-28
Contact: Hu Yong

摘要/Abstract

摘要：

表格存储的BRDF(bidirectional reflectance distribution function)可高逼真再现物体表面外观，但其高维特性和单张平面图包含的反射信息少、差异小的特点，使得现有方法在估计表格存储的BRDF时通常需使用复杂设备或捕获多张图像。针对该问题，提出一种结合神经网络和奇异值分解的单张图材质重建方法。引入奇异值分解将材质压缩到低维空间，将估计材质反射属性的过程简化为估计材质基的权重向量；结合深度卷积神经网络，建立单张平面图和权重向量之间的映射关系，设计了抑制高动态数值变化的对数相对损失函数；为该网络模型的训练构建一个平面样本数据集；提出拆分数据训练方法来消除环状伪影。实验结果表明，该方法在合成数据和真实数据上均优于现有方法。

关键词: 表格存储的BRDF, 神经网络, 奇异值分解, 高动态范围, 材质重建

Abstract:

The tabulated BRDFs (bidirectional reflectance distribution function) can realistically reproduce the surface appearance of objects. However, due to their high-dimensional characteristics and the fact that a single planar image contains limited reflectance information and small differences, methods for estimating tabulated BRDFs typically require complex equipment or the capture of multiple images. To address this issue, a method is proposed for reconstructing material properties from a single image by combining neural networks with singular value decomposition. The singular value decomposition is introduced to compress the material into a lower-dimensional space. The task of solving the tabulated BRDFs is simplified to estimating the weight vectors of material bases. By combining a deep convolutional neural network, a mapping relationship between the single planar image and the weight vector is established. A logarithmic relative loss function is designed to suppress high dynamic range within the material. A planar sample dataset is constructed to train the network model. A two-sided training method is proposed to avoid ringing at the highlights. Experiments results on both synthetic and real data show that the proposed method outperforms previous methods.

Key words: tabulated BRDF, neural networks, singular value decomposition, high dynamic range, material reconstruction

中图分类号:

TP391.9

李志强,沈旭昆,胡勇等 . 结合神经网络和奇异值分解的单图像材质重建[J]. 系统仿真学报, 2026, 38(1): 189-199.

Li Zhiqiang,Shen Xukun,Hu Yong,et al . Material Reconstruction from Single Image Combining Neural Networks with Singular Value Decomposition[J]. Journal of System Simulation, 2026, 38(1): 189-199.

图/表 8

图1

图2

图3

图4

图5

图6

表1

图7

参考文献 30

[1]	Dana K J, van Ginneken Bram, Nayar S K, et al. Reflectance and Texture of Real-world Surfaces[J]. ACM Transactions on Graphics, 1999, 18(1): 1-34.
[2]	Lawrence J, Ben-Artzi A, Decoro C, et al. Inverse Shade Trees for Non-parametric Material Representation and Editing[J]. ACM Transactions on Graphics, 2006, 25(3): 735-745.
[3]	Matusik W. A Data-driven Reflectance Model[D]. Cambridge: Massachusetts Institute of Technology, 2003.
[4]	Jannik Boll Nielsen, Jensen H W, Ramamoorthi R. On Optimal, Minimal BRDF Sampling for Reflectance Acquisition[J]. ACM Transactions on Graphics, 2015, 34(6): 186.
[5]	Xu Zexiang, Jannik Boll Nielsen, Yu Jiyang, et al. Minimal BRDF Sampling for Two-shot Near-field Reflectance Acquisition[J]. ACM Transactions on Graphics, 2016, 35(6): 188.
[6]	Asselin Louis-Philippe, Laurendeau Denis, Lalonde Jean-François. Deep SVBRDF Estimation on Real Materials[C]//2020 International Conference on 3D Vision (3DV). Piscataway: IEEE, 2020: 1157-1166.
[7]	Guo Jie, Lai Shuichang, Tao Chengzhi, et al. Highlight-aware Two-stream Network for Single-image SVBRDF Acquisition[J]. ACM Transactions on Graphics, 2021, 40(4): 123.
[8]	Ye Wenjie, Li Xiao, Dong Yue, et al. Single Image Surface Appearance Modeling with Self-augmented CNNs and Inexact Supervision[J]. Computer Graphics Forum, 2018, 37(7): 201-211.
[9]	Chen Zhe, Nobuhara Shohei, Nishino Ko. Invertible Neural BRDF for Object Inverse Rendering[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022, 44(12): 9380-9395.
[10]	Santo Hiroaki, Samejima Masaki, Sugano Yusuke, et al. Deep Photometric Stereo Networks for Determining Surface Normal and Reflectances[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022, 44(1): 114-128.
[11]	Ren Peiran, Wang Jiaping, Snyder J, et al. Pocket Reflectometry[J]. ACM Transactions on Graphics, 2011, 30(4): 45.
[12]	Lin Y, Peers P, Ghosh A. On-site Example-based Material Appearance Acquisition[J]. Computer Graphics Forum, 2019, 38(4): 15-25.
[13]	Hui Zhuo, Sunkavalli K, Lee J Y, et al. Reflectance Capture Using Univariate Sampling of BRDFs[C]//2017 IEEE International Conference on Computer Vision (ICCV). Piscataway: IEEE, 2017: 5372-5380.
[14]	Deschaintre Valentin, Aittala Miika, Durand Fredo, et al. Single-IMAGE SVBRDF Capture with a Rendering-aware Deep Network[J]. ACM Transactions on Graphics, 2018, 37(4): 128.
[15]	Li Zhengqin, Sunkavalli K, Chandraker M. Materials for Masses: SVBRDF Acquisition with a Single Mobile Phone Image[C]//Computer Vision – ECCV 2018. Cham: Springer International Publishing, 2018: 74-90.
[16]	Wen Tao, Wang Beibei, Zhang Lei, et al. SVBRDF Recovery from a Single Image with Highlights Using a Pre-trained Generative Adversarial Network[J]. Computer Graphics Forum, 2022, 41(6): 110-123.
[17]	Zhao Yezi, Wang Beibei, Xu Yanning, et al. Joint SVBRDF Recovery and Synthesis from a Single Image Using an Unsupervised Generative Adversarial Network[C]//Eurographics Symposium on Rendering - DL-only Track. Geneva: The Eurographics Association, 2020: 53-66.
[18]	Deschaintre Valentin, Aittala M, Durand F, et al. Flexible SVBRDF Capture with a Multi-image Deep Network[J]. Computer Graphics Forum, 2019, 38(4): 1-13.
[19]	Gao Duan, Li Xiao, Dong Yue, et al. Deep Inverse Rendering for High-resolution SVBRDF Estimation from an Arbitrary Number of Images[J]. ACM Transactions on Graphics, 2019, 38(4): 134.
[20]	Guo Yu, Smith C, Hašan Miloš, et al. MaterialGAN: Reflectance Capture Using a Generative SVBRDF Model[J]. ACM Transactions on Graphics, 2020, 39(6): 254.
[21]	Li Zhiqiang, Shen Xukun, Hu Yong, et al. High-resolution SVBRDF Estimation Based on Deep Inverse Rendering from Two-shot Images[J]. The Visual Computer, 2023, 39(10): 4609-4622.
[22]	Deschaintre Valentin, Drettakis G, Bousseau A. Guided Fine-tuning for Large-scale Material Transfer[J]. Computer Graphics Forum, 2020, 39(4): 91-105.
[23]	Zhou Xilong, Kalantari N K. Adversarial Single-image SVBRDF Estimation with Hybrid Training[J]. Computer Graphics Forum, 2021, 40(2): 315-325.
[24]	Goodfellow I, Pouget-Abadie Jean, Mirza Mehdi, et al. Generative Adversarial Networks[J]. Communications of the ACM, 2020, 63(11): 139-144.
[25]	Zhang Xiuming, Srinivasan P P, Deng Boyang, et al. NeRFactor: Neural Factorization of Shape and Reflectance Under an Unknown Illumination[J]. ACM Transactions on Graphics, 2021, 40(6): 237.
[26]	Zhou Shuyao, Zhu Tianqian, Shi Kanle, et al. Review of Light Field Technologies[J]. Visual Computing for Industry, Biomedicine, and Art, 2021, 4(1): 29.
[27]	Matusik W, Pfister H, Brand M, et al. Efficient Isotropic BRDF Measurement[C]//Proceedings of the 14th Eurographics Workshop on Rendering. Goslar: Eurographics Association, 2003: 241-247.
[28]	He Kaiming, Zhang Xiangyu, Ren Shaoqing, et al. Deep Residual Learning for Image Recognition[C]//2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2016: 770-778.
[29]	Hu Bingyang, Guo Jie, Chen Yanjun, et al. DeepBRDF: A Deep Representation for Manipulating Measured BRDF[J]. Computer Graphics Forum, 2020, 39(2): 157-166.
[30]	Zhang R, Isola P, Efros A A, et al. The Unreasonable Effectiveness of Deep Features as a Perceptual Metric[C]//2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2018: 586-595.

方法	RMSE(BRDF)	RMSE(渲染球体)	LPIPS(渲染球体)
本文	0.169 86	0.018 82	0.014 00
文献[4]1次采样	0.542 77	0.118 63	0.118 29
文献[4]2次采样	0.391 95	0.036 14	0.031 63
文献[4]5次采样	0.207 61	0.035 43	0.024 16
文献[14]		0.056 38	0.045 14
文献[23]		0.055 48	0.049 34

结合神经网络和奇异值分解的单图像材质重建

Material Reconstruction from Single Image Combining Neural Networks with Singular Value Decomposition

RichHTML

PDF (PC)

可视化

摘要/Abstract

引用本文

使用本文

图/表 8

参考文献 30

相关文章 15

编辑推荐

Metrics

本文评价

[1]	魏呈彪, 赵涛岩, 曹江涛, 李平. 深度模糊神经网络的设计和预测[J]. 系统仿真学报, 2025, 37(9): 2200-2210.
[2]	吕金虎, 蒋弘毅, 刘德元, 谭少林. 基于图神经网络的复杂系统建模与仿真[J]. 系统仿真学报, 2025, 37(7): 1624-1638.
[3]	王子怡, 张凯, 钱殿伟, 刘玉贞. 一种基于DRL的分布式装备体系优选方法[J]. 系统仿真学报, 2025, 37(6): 1565-1573.
[4]	张鹏, 冯柯, 宫建成, 杨小强, 申金星. 基于RBF神经网络的防空导弹武器系统作战效能评估[J]. 系统仿真学报, 2025, 37(2): 529-540.
[5]	沈嘉玮, 才大业, 杨国青, 吕攀, 李红. 大规模脉冲神经网络动态加载仿真方法[J]. 系统仿真学报, 2025, 37(2): 541-550.
[6]	李孝斌, 胡冰, 尹超, 李波, 马军. 基于时空图卷积的汽车配件供应链需求预测与仿真分析[J]. 系统仿真学报, 2025, 37(12): 3060-3074.
[7]	张文康, 孙霄峰, 钟一平, 尹勇. 基于图神经网络的船舶液舱晃荡数值仿真[J]. 系统仿真学报, 2025, 37(12): 3087-3099.
[8]	许璟琳, 陈芊茹, 彭阳, 余芳强. 基于GRU-模拟退火的公共建筑自校准客流仿真及空间优化研究[J]. 系统仿真学报, 2025, 37(11): 2826-2838.
[9]	田元兴, 韩则胤, 王宁, 苏宝定, 向未林. 基于时空特性的齿轮箱故障预警系统孪生建模[J]. 系统仿真学报, 2025, 37(11): 2867-2876.
[10]	张志利, 刘瑾, 周召发, 梁哲, 张云昊. 基于ISCSO-BP神经网络模型的光纤陀螺温度补偿技术研究[J]. 系统仿真学报, 2025, 37(11): 2904-2917.
[11]	李博宁, 陈明, 齐舒畅, 孟浩然, 王磊. 基于扰动观测器和指令滤波的四旋翼无人机姿态控制[J]. 系统仿真学报, 2025, 37(11): 2793-2803.
[12]	黄德启, 涂亚婷, 张振华, 郭鑫. 基于MLP与改进GCN-TD3的交通信号控制建模与仿真[J]. 系统仿真学报, 2025, 37(10): 2568-2577.
[13]	顾皓, 王佳宇, 熊伟丽. 双流框架下的改进Transformer软测量建模[J]. 系统仿真学报, 2025, 37(10): 2594-2604.
[14]	蒋文彬, 曹余庆, 谢莉, 杨慧中. 基于SALR网络的双率采样非线性系统鲁棒辨识[J]. 系统仿真学报, 2025, 37(10): 2652-2661.
[15]	刘晓德, 郭宇飞, 陈元培, 周洁, 张瑀涵, 彭玮航, 马喆. 基于脉冲强化学习的连续运动控制仿真与优化[J]. 系统仿真学报, 2025, 37(10): 2662-2671.