A MULTIPLE-PROCESSOR ARCHITECTURE FOR IMAGE PROCESSING

This thesis is concerned with the design and analysis of a parallel computer architecture for image processing. The proposed architecture is capable of supporting at high efficiency a wide range of image processing tasks that are diverse in their computational structures by means of dynamic algorithm-architecture matching. This architectural versatility is achieved primarily by making use of a novel storage scheme in conjunction with a unique multistage interconnection network which allow: (1) various N -partitions of an image; (2) parallel access of any part of an N -partition of the image without memory conflict; (3) placement of the elements of a given part on a processing element array in useful orders for subsequent highly parallel processing. Fundamental properties of the architecture are given, and some implementation considerations are provided. The use of the architecture for the efficient parallel implementation of three fundamental classes of computationally intensive image processing tasks is investigated. They are local neighborhood operations, global neighborhood operations, and global transforms. It is shown that although they are diverse in their computational structures, the architecture allows them to achieve a high degree of parallelism in the computation.