Real-Time Time-Warped Multiscale Signal Processing for Scientific Visualization

This thesis considers the problem of visualizing simulations of phenomenon which span large ranges of spatial scales. These datasets tend to be extremely large presenting challenges both to human comprehension and high-performance computing. The main problems considered are how to effectively represent scale and how to efficiently compute and visualize multiscale representations for large, real-time datasets. Time-warped signal processing techniques are shown to be useful for formulating a localized notion of scale. In this case, we use time-warping in order to adapt the standard Fourier basis to local properties of the signal, giving the advantage of being localized in the frequency spectrum as compared with the standard linear notions of scale. Time-warping is also shown to have theoretical advantages in terms of signal reconstruction quality and random noise removal. In practice, these advantages are shown to only hold under certain conditions. It is then shown in the thesis how convolution-based reconstruction techniques can be mapped onto graphics processing units (GPUs) for high-performance implementation of a multiscale molecular visualization framework. We show how the same technique can likely be used for time-warped multiscale reconstruction.