Optimal image algorithms on an orthogonally-connected memory-based architecture

Processor-time optimal algorithms are presented for several image and vision problems. A parallel architecture which combines an orthogonally accessed memory with a linear array structure is used. The organization has p processors and a memory of size O(n²) locations. The number of processors p can vary over the range [1, n^3/2] while providing optimal speedup for several problems in image analysis and vision. Such problems include labeling connected regions, computing minimum convex containers of regions, and computing nearest neighbors of pixels and regions. Optimal algorithms are presented for histogramming and computing the Hough transform. Such problems arise in medium-level vision and require global operations or dense data movement. It is shown that for these types of problems, the proposed organization is superior to the mesh and pyramid organizations.