基于RBF神经网络的全局光照实时绘制  

Real-Time Global Illumination Based on RBF Neural Network

作　　者：刘晓芸 姚承宗[1] 曾晓勤[1] LIU Xiao-yun;YAO Cheng-zong;ZENG Xiao-qin(School of Computer and Information,Hohai University,Nanjing Jiangsu 211100,China)

机构地区：[1]河海大学计算机与信息学院,江苏南京211100

出　　处：《计算机仿真》2021年第9期424-428,共5页Computer Simulation

基　　金：国家自然科学基金(60971088)。

摘　　要：提出一种采用径向基函数RBF神经网络对间接光照进行拟合的方法,以实现对动态视点场景的全局光照实时绘制。使用离线渲染的方式进行预计算,获得训练数据,通过减聚类的方法确定RBF的中心,采用有监督的学习方法进行训练,拟合间接光照中的视点位置、光源位置、场景中物体表面法向量等与间接光照值之间的非线性关系,以代替传统的全局光照模型完成间接光照的计算,避免了光线的多次求交。实验结果表明,上述方法能有效减少冗余数据,以较快的速度收敛,并且很好地拟合离线渲染的结果,完成全局光照实时绘制。In this paper,an approach is proposed for fast rendering global illumination in scene of dynamic viewpoint,which uses the Radial Basis Function Neural Network to simulate the indirect illumination.The training data for the neural network were precomputed by offline render and the subtractive clustering method was used to set the radial basis function center.In order to find the nonlinear coherence between the attributes such as viewing direction,light source,surface normal,supervised learning was chosen instead of the classic ray-tracing algorithm.Our approach is effective by avoiding the computation of inter-reflection.Experimental results show that the approach is capable of reducing the redundant training data,evaluating fast and rendering well.

关 键 词：全局光照 径向基函数神经网络 减聚类 机器学习 间接光照 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]