Learning parameters for object recognition by the self-organizing Hopfield network

Recently, the authors introduced a novel programming strategy to generate homomorphic graph matching using the Hopfield network and a Lyapunov indirect method based constraint parameter learning scheme. In this paper, an augmented weighted model attributed relational graph (WARG) representation scheme and learning schemes to estimate parameters used in the WARG representation and compatibility measure function are presented to improve further the homomorphic graph matching strategy. The representation scheme incorporates a distinct weighting factor and tolerance parameter for every model attribute. To estimate the parameters in a simplified form of the model WARG representation, learning schemes are presented. The computation of weighting factors is formulated as an optimization problem and solved using the quadratic programming algorithm. The formulation implicitly evaluates ambiguity, robustness and discriminatory power of the relational attributes chosen for graph matching and assigns weighting factors appropriately to the chosen attributes. Experimental results are presented to demonstrate that the parameter learning schemes are essential when the models have intra-model ambiguity and the optimal set of parameters always generate a better mapping.