Monte Carlo (MC) path tracing causes a lot of noise on the rendered image at a low samples per pixel. Recently, with the help of inexpensive auxiliary buffers and the generative adversarial network (GAN), deep learning-based denoising MC rendering methods have been able to generate noise-free images with high perceptual quality in seconds. In this paper, we propose a novel GAN structure for denoising Monte Carlo renderings, called dual residual connection GAN. Our key insight is that the dual residual connections can improve the chance of the optimal feature selection and implicitly increase the number of potential interactions between modules. We also propose a multi-scale auxiliary features extraction method, aiming to make full use of the rich geometry and texture information of auxiliary buffers. Moreover, we adopt a spatial-adaptive block with the deformable convolution to help the network adapt to the variance in spatial texture and edge features. Compared with the state-of-the-art methods, our network has fewer parameters and less inference time, and the results surpass the previous in terms of visual effects and quantitative metrics.

Monte Carlo (MC) 路径跟踪会在每个像素的样本较低时在渲染图像上产生大量噪声。最近，在廉价的辅助缓冲区和生成对抗网络 (GAN) 的帮助下，基于深度学习的去噪 MC 渲染方法已经能够在几秒钟内生成具有高感知质量的无噪声图像。在本文中，我们提出了一种新的 GAN 结构，用于对 Monte Carlo 渲染进行去噪，称为双残差连接 GAN。我们的主要见解是双残差连接可以提高最佳特征选择的机会，并隐式增加模块之间潜在交互的数量。我们还提出了一种多尺度辅助特征提取方法，旨在充分利用辅助缓冲区丰富的几何和纹理信息。而且，我们采用具有可变形卷积的空间自适应块来帮助网络适应空间纹理和边缘特征的变化。与最先进的方法相比，我们的网络具有更少的参数和更少的推理时间，并且结果在视觉效果和量化指标方面都超过了之前的方法。