A Machine Learning Approach for Filtering Monte Carlo Noise

摘要

The most successful approaches for filtering Monte Carlo noise usernfeature-based filters (e.g., cross-bilateral and cross non-local meansrnfilters) that exploit additional scene features such as world positionsrnand shading normals. However, their main challenge is finding thernoptimal weights for each feature in the filter to reduce noise butrnpreserve scene detail. In this paper, we observe there is a complexrnrelationship between the noisy scene data and the ideal filter parameters,rnand propose to learn this relationship using a nonlinearrnregression model. To do this, we use a multilayer perceptron neuralrnnetwork and combine it with a matching filter during both trainingrnand testing. To use our framework, we first train it in an offline processrnon a set of noisy images of scenes with a variety of distributedrneffects. Then at run-time, the trained network can be used to drivernthe filter parameters for new scenes to produce filtered images thatrnapproximate the ground truth. We demonstrate that our trained networkrncan generate filtered images in only a few seconds that are superiorrnto previous approaches on a wide range of distributed effectsrnsuch as depth of field, motion blur, area lighting, glossy reflections,rnand global illumination.