Denoising Monte Carlo Images with Broadly-Trained and Scene-Specific Neural Network Parameters

Abstract

Computer graphics are widely used nowadays to render a variety of digital scenes. A rendering technique of particular interest to us is path tracing. Path tracing can generate photorealistic images and has recently grown in popularity. However, path tracing is a computationally intensive and iterative rendering technique, and generating an image using path tracing requires a large amount of iterations and thus time. One way to vastly decrease the amount of time required is to use a low-iteration image and to denoise that image using a denoising filter based on artificial intelligence, in our case a convolutional neural network. Such a neural network must be trained, and this can be done on a large variety of training data for use on many different scenes, or on a small set of training data for use on one specific scene. The purpose of this thesis is to show the decrease in denoising error on images generated from some scene, when comparing scene-specific training data to broad training data. We do this by using a convolutional neural network with two different sets of network parameters per scene. We experiment with a set of scene-specific network parameters for each scene, and a set of broadly-trained network parameters for all scenes. Our experiments show that scene-specific network parameters can lead to a notable decrease in denoising error under certain circumstances. Further research may however be required to form a more concrete result.