Self-organizing maps and full GPU parallel approach to graph matching

Abstract

Graph matching is an essential problem in computer science and communications. It can be applied to a variety of issues such as artificial intelligence, computer vision, and communication systems. In this paper, we propose a new Graphics Processing Unit framework written in CUDA C++ specifically dedicated to geometric graph matching but providing new parallel algorithms, with low computational complexity, as the self-organizing map in the plane, and a distributed local search method. Unlike state-of-the-art graph matching algorithms, available from Matlab platform, that most often need at least O(N2) memory size, with N the problem size, our proposals only require O(N) space and allows massively parallel execution. These parallel algorithms are evaluated and compared to the state-of-the-art methods available for graph matching and following the same experimental protocol.