LANDMARK-BASED SHAPE-RECOGNITION BY A MODIFIED HOPFIELD NEURAL-NETWORK

A new method to recognize partially visible two-dimensional objects by means of a modified Hopfield neural network is introduced. Each object is represented by a set of ''landmarks'', and thus such an approach is referred to as landmark-based shape recognition. The landmarks of an object are points of interest relative to the object that have important shape attributes. Given a scene consisting of partially visible objects, a model object in the scene is determined by how well the model landmarks are matched to those in the scene. A local shape measure, known as the sphericity, is used to measure the similarity between two landmarks. The hypothesis of a model object in a scene is made by matching the model landmarks with the scene landmarks. The landmark matching task is performed by a modified Hopfield neural network. The location of the model in the scene is estimated by a least squares fit among the matched landmarks. The hypothesis is finally verified by a heuristic measure. The convergence of the modified Hopfield neural network is proven, and the robustness of the approach has been experimentally demonstrated.