Parallel coordination of image operators: model, algorithm and performance

To obtain machine vision algorithms which are robust in the face of variations in image lighting, arrangements of objects, viewing parameters, etc., it is helpful to build into the algorithms an adaptive control mechanism such as a state-space search procedure. Such a procedure dynamically determines an optimal sequence of image processing operators to classify an image or to put its parts into correspondence with a model or set of models. One benefit of structuring the vision algorithm as a state-space search is that a multiplicity of paths toward goal nodes in the state space can be explored concurrently. In this paper, we identify several types of parallelism that may be exploited in vision algorithms based on state-space search. We present a new method, the ‘V ∗ algorithm’, which, unlike earlier parallel search algorithms, generates the successors of a state in parallel. In machine vision, this part of the search process is very expensive, and thus V ∗ permits substantial speedup. An experimental evaluation of V ∗ is presented which is based on a simulation of a character recognition algorithm; the simulation runs on a Sequent Balance 21000 using 16 processors.