Object parts matching using Hopfield neural networks

Abstract

An optimization approach is used to solve the Cyclic Ordered Assignment (COA) problem which occurs when matching 2D object parts for recognition. The solution space for the COA problem becomes very large when partially occluded objects are considered. By associating the solutions of the COA problem with the local minima of the energy function for a 2D binary Hopfield network, a network is presented which can solve the problem by converging from an initial state to a local minima. The initial state of the network is an array representing the probabilities of matches between the corresponding parts of an unknown object and a known template object. By taking advantage of the computational power and parallel processing of the network we can arrive at a fast, accurate solution for each input state presented to the network.