Optimising object recognition parameters using a parallel multiobjective genetic algorithm

This paper describes application of a multiobjective genetic algorithm (MOGA) to optimise the selection of parameters for an object recognition scheme. The MOGA applied uses Pareto-ranking as a means of comparing individuals over multiple objectives. In order to prevent premature convergence heuristics were added to the algorithm to encourage speciation. The population consisted of sub-populations, whose members were able to migrate to and other sub-population, thus following the `island´ population model. Prior to this work the pairwise geometric histogram (PGH) object recognition paradigm required the user to manually select histogram parameters - a process involving some degree of experience with the recognition scheme. Here, through the application of a MOGA we optimise and consequently automate parameter selection. The overall result of the algorithm is to select PGH parameters giving a more compact efficient histogram representation