Document Type
Thesis
Date of Award
5-31-1991
Degree Name
Master of Science in Electrical Engineering - (M.S.)
Department
Electrical Engineering
First Advisor
Sotirios Ziavras
Second Advisor
John D. Carpinelli
Third Advisor
Edwin Hou
Abstract
Hypercubes and meshes are of the most successful parallel processing structures and are widely used in image processing applications. The mesh is a natural multiprocessor architecture for image processing because an N x N image can be directly mapped onto a mesh of N x N PEs (processing elements), holding each PE responsible for one pixel, with neighboring PEs responsible for neighboring pixels. Thus, local operations can be performed on the image, or local image features can be detected in parallel, with each processor simultaneously accessing its neighbor and computing the appropriate function for its neighborhood. Meshes and several other structures can be easily mapped onto hypercubes making them a very reasonable choice for a variety of image processing applications. In addition, hypercubes also support global communications at a very reasonable cost. The separation of objects from their background is a major problem in pattern recognition and scene analysis. It finds applications in many fields, ranging from optical character recognition to biomedical image processing. Several algorithms already have been proposed to solve this problem. An important subset of these algorithms are implemented here on the mesh connected massively parallel BLITZEN supercomputer. In addition, another algorithm is proposed and simulated for incomplete hypercube structures. Finally, a comparative analysis is carried out based on the results obtained from the implementations on BLITZEN and the presented simulations.
Recommended Citation
Shahid, Imran, "Connected component labelling on a pair of parallel systems" (1991). Theses. 2613.
https://digitalcommons.njit.edu/theses/2613
