Interactive Three-Dimensional Image Analysis and Visualization Using Graphics Hardware

Image analysis is a process of extracting useful information from input images using various image processing techniques, such as denoising, enhancement, feature detection, and segmentation. State-of-the-art techniques employ sophisticated and time-consuming nonlinear processes to capture subtle and detailed information from images that may be overlooked by classical linear methods. In addition, with the advent of the recent advance in imaging technology, high resolution images of several hundreds gigabytes become common these days. Such recent trends in image processing require computationally expensive processes and make the interactive image analysis challenging. In this dissertation I propose novel three-dimensional nonlinear image processing algorithms and their efficient implementations on modern graphics processors for interactive image analysis, specifically targeting the geoscience and medical applications. Over the past decade, the graphics processing unit (CPU) has evolved into a powerful computing platform for high performance computation. Among many applications, image processing can be a good candidate to achieve a significant speed-up using the GPU because it is inherently a data parallel process. The main motivation of the proposed dissertation research is to accelerate time-consuming image analysis applications using inexpensive commodity CPUs. This dissertation consists of three parts. First, an interactive 3D seismic fault detection system is proposed. I have developed a structure-oriented directional anisotropic diffusion filter and a variance-based fault detection algorithm, and introduced their efficient implementations on the GPU. Second, I propose a fast parallel algorithm for weighted distance computation using a Hamilton-Jacobi framework. The proposed algorithm is based on a label-correcting scheme for scalability on single instruction multiple data (SIMD) parallel architectures. Finally, I propose a probabilistic framework for user-assisted 3D elongated structure segmentation, and show the acceleration techniques using the CPU implementation of nonrigid image registration and probability computation. A number of synthetic and real-world 3D datasets, such as seismic, CT, and DT-MRI volumes, are used to assess the validity and effectiveness of the proposed methods.