交互式蒙特卡洛全局光照超分辨率重建  

Interactive Monte Carlo global illumination super-resolution reconstruction

作　　者：王静雯 梁晓[1] 王妮婷 欧阳娇 WANG Jing-wen;LIANG Xiao;WANG Ni-ting;OU Yang-jiao(School of Computer Science,Southwest Petroleum University,Chengdu 610500,China)

机构地区：[1]西南石油大学计算机科学学院,四川成都610500

出　　处：《计算机工程与设计》2023年第4期1166-1173,共8页Computer Engineering and Design

基　　金：国家自然科学基金项目(61873218);四川省科技厅应用基础研究基金项目(18YYJC1147);西南石油大学创新基金项目(642)。

摘　　要：针对现有全局光照图像的重建效果模糊且耗时较长问题,提出一种基于生成对抗的交互式图像超分辨率重建网络,使用低分辨率的高频噪声推理高分辨率的全局光照图像。将G-buffer中的图形特征作为辅助输入,利用其低噪与相似性,为网络增加图像高频信息;为避免多模态输入下产生冗余特征,利用注意力机制抑制辅助输入中的无效信息;建立混合感知损失函数,促进产生自然的图像风格。实验结果表明,在4 spp的低分辨率噪声输入下,能够在交互式帧率下获得4倍超分辨率全局光照图像,支持阴影、间接漫/镜面反射等光照效果。To address the problem of blurred and time-consuming reconstruction of existing global illumination images,an inte-ractive image super-resolution reconstruction network based on generative adversarial was proposed to infer high-resolution global illumination images using low-resolution high-frequency noise.The graphical features in the G-buffer were used as auxiliary inputs to add high-frequency information of the image to the network by exploiting its low noise and similarity.To avoid redundant features under multimodal input,an attention mechanism was used to suppress invalid information in the auxiliary input.A hybrid perceptual loss function was established to facilitate the generation of natural image styles.Experimental results show that 4x super-resolution global illumination images can be obtained at interactive frame rates with low-resolution noise input of 4 spp,and support illumination effects such as shadows and indirect diffuse/specular reflections.

关 键 词：全局光照 超分辨率重建 生成对抗网络 注意力机制 交互式绘制 蒙特卡洛降噪 辅助特征 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]