Date of Award


Document Type


Degree Name

Doctor of Philosophy (PhD)


Computer Science

First Advisor

S. Sitharama Iyengar


Parallel processing has been considered as the key to build computer systems of the future and has become a mainstream subject in Computer Science. Computer Vision applications are computationally intensive that require parallel approaches to exploit the intrinsic parallelism. This research addresses this problem for low-level and intermediate-level vision problems. The contributions of this dissertation are a unified scheme based on probabilistic relaxation labeling that captures localities of image data and the ability of using this scheme to develop efficient parallel algorithms for Computer Vision problems. We begin with investigating the problem of skeletonization. The technique of pattern match that exhausts all the possible interaction patterns between a pixel and its neighboring pixels captures the locality of this problem, and leads to an efficient One-pass Parallel Asymmetric Thinning Algorithm (OPATA$\sb8).$ The use of 8-distance in this algorithm, or chessboard distance, not only improves the quality of the resulting skeletons, but also improves the efficiency of the computation. This new algorithm plays an important role in a hierarchical route planning system to extract high level typological information of cross-country mobility maps which greatly speeds up the route searching over large areas. We generalize the neighborhood interaction description method to include more complicated applications such as edge detection and image restoration. The proposed probabilistic relaxation labeling scheme exploit parallelism by discovering local interactions in neighboring areas and by describing them effectively. The proposed scheme consists of a transformation function and a dictionary construction method. The non-linear transformation function is derived from Markov Random Field theory. It efficiently combines evidences from neighborhood interactions. The dictionary construction method provides an efficient way to encode these localities. A case study applies the scheme to the problem of edge detection. The relaxation step of this edge-detection algorithm greatly reduces noise effects, gets better edge localization such as line ends and corners, and plays a crucial rule in refining edge outputs. The experiments on both synthetic and natural images show that our algorithm converges quickly, and is robust in noisy environment.