Volume Visualization Using Principal Component Analysis

Abstract

In this paper, we investigate the use of Principal Component Analysis (PCA) for image-based volume visualization. Firstly we compute a high-dimensional eigenspace using training images, pre-rendered using a standard ray-caster, from a spherically distributed range of camera positions. Then, our system is able to synthesize arbitrary views of the dataset with minimal computation at runtime. We propose a perceptually-adaptive technique to minimize data size and computational complexity whilst preserving perceptual quality of the visualization, in comparison to corresponding ray-cast images. Results indicate that PCA is able to sufficiently learn the full view-independent volumetric model through a finite number of training images and generalize the computed eigenspace to produce high quality images from arbitrary viewpoints, on demand. The approach has potential application in client-server volume visualization or where results of a computationally-complex 3D imaging process need to be interactively visualized on a display device of limited specification.