基于卷积神经网络的实时环境光遮蔽计算  

Convolution neural network based real-time ambient occlusion computing

作　　者：郭雨潇 陈雷霆[1,2,3] Guo Yuxiao;Chen Leiting(School of Computer Science & Engineering,University of Electronic Science & Technology of China,Chengdu 611731,China;Institute of Electronic & Information Engineering of Dongguan,University of Electronic Science & Technology of China,Dongguan Guangdong 523808,China;Government Affairs Service Center of Guangdong Provincial Health Department,Guangzhou 510060,China)

机构地区：[1]电子科技大学计算机科学与工程学院,成都611731 [2]电子科技大学广东电子信息研究院,广东东莞523808 [3]广东省卫生厅政务服务中心,广州510060

出　　处：《计算机应用研究》2019年第6期1883-1886,1896,共5页Application Research of Computers

基　　金：国家“863”计划资助项目(2015AA016010);广东省应用型科技研发专项资金资助项目(2015B010131002);东莞市重大科技项目(2015215102)

摘　　要：已有的基于屏幕空间的环境光遮蔽计算方法虽然保证了计算的实时性,但是存在估计失真以及细节丢失等难以解决的问题。针对该问题,提出了一种结合低频光线追踪采样以及蒙特卡罗去噪的算法框架对环境光遮蔽进行实时计算。为了解决传统蒙特卡罗去噪算法无法实时处理的问题,提出了一种基于卷积神经网络的蒙特卡罗去噪算法,并针对问题对网络结构进行了改进和优化。实验证明基于卷积神经网络的方法能够对环境光遮蔽的去噪问题进行准确的处理,同时对卷积网络的改进在保持精度的基础上显著地提高了计算效率。对比实验显示了该算法在保持与高频采样光线追踪算法相近效果的前提下可达到与基于屏幕空间环境光遮蔽计算方法相近的每秒数百帧的计算效率。Although existing screen-space solutions can provide real-time performance, they cannot generate accurate results and preserve details. This paper proposed a new solution based on Monte-Carlo ray tracing. To achieve real-time performance, this method sampled the visibility of each pixel with sparse rays and then denoised the ray tracing results with a convolutional neural network (CNN). To meet the real-time requirement of ambient occlusion computing, the algorithm further improved and optimized the network structure. Experimental results show that proposed method not only achieves the real-time performance, but also produces more accurate and detailed results.

关 键 词：环境光遮蔽 蒙特卡罗采样去噪 卷积神经网络 屏幕空间 自动编码器 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]