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。
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.
作者
李志强
沈旭昆
胡勇
周雪杨
陈弈帆
Li Zhiqiang;Shen Xukun;Hu Yong;Zhou Xueyang;Chen Yifan(National Engineering Laboratory for Modeling and Emulation in E-Government,Harbin Engineering University,Harbin 150001,China;State Key Laboratory of Virtual Reality Technology and Systems,Beihang University,Beijing 100191,China)
出处
《系统仿真学报》
北大核心
2026年第1期189-199,共11页
Journal of System Simulation
基金
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)。
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
tabulated BRDF
neural networks
singular value decomposition
high dynamic range
material reconstruction
