Survey of parallel algorithms for structural pattern matching

Matching is an important part of a model-based object recognition system. Matching is a difficult task, for a number of reasons. First, in a number of recognition systems matching is formulated as a combinatorial problem with exponential worst-case complexity. Thus, heuristics are needed to reduce the complexity by pruning the search space. Second, images do not present perfect data: noise and occlusion greatly complicate the task. Finally, even at moderate image resolutions the amount of data to be handled is such that this task cannot be done in real-time on supercomputers. Although no existing visual system can solve the general recognition problem, some existing approaches have obtained acceptable results for limited domains or simple scenes. Much less work has been done on parallel matching, despite the great need for speeding up the process. Parallel algorithms have often to be designed from scratch, and the recognition problem itself often requires reformulation since many of the proposed sequential algorithms do not lend themselves naturally to efficient parallel implementations. In this paper, we survey some of the existing parallel matching algorithms for 2D and 3D objects. Some of these algorithms have been implemented on SIMD architectures such as the Connection Machine or MasPar, or MIMD machines such as the Intel Touchstone Delta; other algorithms have been developed for the PRAM model of computation