Path Tracing Denoising based on SURE Adaptive Sampling and Neural Network

Qiwei Xing, Chunyi Chen
2020 IEEE Access  
A novel reconstruction algorithm is presented to address the noise artifacts of path tracing. SURE (Stein's unbiased risk estimator) is adopted to estimate the noise level per pixel that guides adaptive sampling process. Modified MLPs (multilayer perceptron) network is used to predict the optimal reconstruction parameters. In sampling stage, coarse samples are firstly generated. Then each noise level is estimated with SURE. Additional samples are distributed to the pixels with high noise level.
more » ... Next, we extract a few features from the results of adaptive sampling used for the subsequent reconstruction stage. In reconstruction stage, modified MLPs network is adopted to model a complex relationship between extracted features and optimal reconstruction parameters. An anisotropic filter is used to reconstruct the final images with the parameters predicted by neural networks. Compared to the state-of-the-art methods, experiment results demonstrate that our algorithm performs better than other methods in numerical error and visual image quality. INDEX TERMS Adaptive sampling, SURE estimator, MLPs network, path tracing, denoising. In order to combine the advantages of two kind of approaches, we propose a novel algorithm that leverages the power of neural networks in the following key manners compared to previous denoising methods: • Rather than predicting the pixel color, we utilize the neural network to predict the optimal reconstruction parameters. Then we can make use of a cross-bilateral filters to reconstruct the final noise-free images with these predicted parameters. And the quality of final images is higher. • Compared with those methods which used the neural network to predict the kernel size of the filter, we introduce a set of additional features extracted from the path tracing rendering process, such as the mean and deviation of color in the neighborhood around the pixel, which is computed on each color component, the world position of the hit-points. And we set them as the input of neural network. The more the extracted features, the more accurate relationship between the filter parameters and the features. In other words, we can get the more optimal filter parameters easily.
doi:10.1109/access.2020.2999891 fatcat:l3efvetnefbptez3ejknvucvgy