Convolution: Making Rendered Images Clearer
No Thumbnail Available
Authors
Grosse, Jacob
Date of Issue
2023-04-28
Type
Presentation
Language
Subject Keywords
Other Titles
Abstract
Realistic image rendering is a computationally expensive process that, even when done efficiently, can take hours, days, or even weeks to render a single image. This research is an explorative journey through making this process more efficient with unconventional means. Last semester (Fall 2022) was spent developing a ray-tracer, or an application which renders a scene by simulating light rays realistically. This process is done pixel by pixel for any given image and as such noise can be generated due to the somewhat random nature of light and how it bounces off different surfaces in simulation. The only solution to remove the noise generated from rendering the ray-traced images that was found during this development time was to simply render each pixel several thousand times and average the results, which results in an obscenely long runtime for the application. It is for this reason that we have chosen to explore the possibility of using neural networks, a series of functions that are organized into layers which are tweaked through the process of “training” to achieve a desired result. By training a neural network to take in noisy images and produce de-noised images, we can cut the processing time of producing high quality images significantly. We will quantify the improvement by using a metric we have created that we are calling aggregate pixel difference. This will measure the difference in color values between adjacent pixels and sum them across the whole image for a single value we wish to minimize. We will use this in conjunction with the mean squared error of the color values per pixel between the initial image and the smoothed image. The ratio at which we weight these will determine how much smoothing takes place, as weighting entirely into aggregate pixel difference will result in an image that is just a solid color and weighting entirely into mean squared error will result in the exact same image we put in.