基于注意力机制的蒙特卡罗渲染去噪方法  

A MONTE CARLO RENDERING DENOISING METHODBASED ON ATTENTION MECHANISM

作　　者：何凯 陈纯毅[1] 胡小娟[1] 于海洋[1] He Kai;Chen Chunyi;Hu Xiaojuan;Yu Haiyang(School of Computer Science and Technology,Changchun University of Science and Technology,Changchun 130022,Jilin,China)

机构地区：[1]长春理工大学计算机科学技术学院,吉林长春130022

出　　处：《计算机应用与软件》2023年第11期241-247,共7页Computer Applications and Software

基　　金：国家自然科学基金项目(U19A2063);吉林省科技发展计划项目(20190302113GX);吉林省教育厅“十三五”科学技术研究项目(JJKH20200792KJ,JJKH20200799KJ)。

摘　　要：针对蒙特卡罗渲染在低采样率下绘制图像的噪声问题,提出基于注意力机制的蒙特卡罗噪声去除方法。在模型上,设计去除批归一化层的注意力残差块减少模型的计算量,引入空间注意力和通道注意力关注渲染图像的细节。使用图像的颜色和辅助信息作为网络的输入,加入感知损失函数训练网络。实验结果表明,相比其他蒙特卡罗渲染去噪方法,该方法在去除噪声的同时不仅取得了更高的量化值,而且保留了更多的细节特征。To solve the problem of rendering noise in Monte Carlo synthesized images induced by low sampling rate,a denoising method based on the attention mechanism is proposed.On the model,the design removed the attention residual block of the batch normalization layer to reduce the amount of calculation of the model,and introduced the spatial attention and the channel attention to pay attention to the details of the rendered image.The color and auxiliary information of the image were used as the input of the network,and the perceptual loss function was added to the training network.Experimental results show that compared with other Monte Carlo rendering denoising methods,the proposed method not only gets higher quantization values but also retains more detailed features.

关 键 词：蒙特卡罗渲染 路径追踪 注意力机制 深度学习 图像去噪 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]