Parallelism in Computer Vision: a Review

Abstract

Computer vision tasks require an enormous amount of computation, especially when the data is in image form, demanding high-performance computers for practical, real-time applications. Parallelism appears to be the only economical way to achieve the level of performance required for vision tasks. Researchers in human and machine vision share the belief that massive parallel processing characterizes low-level vision. In this paper we review the various parallel algorithms used in Computer Vision. The problem of visual recognition is divided into three conceptual levels-low-level, intermediate-level and high-level. There are few conceptual difficulties in parallelizing low-level vision. Intermediate-level vision is relatively difficult. Most algorithms in these two levels have been parallelized. However, not much work has been done in high-level vision. We present a survey of algorithms within each of the three levels.