An analog-Hopfield neural network for recognition of partially visible 2-D objects

Abstract

An application of neural networks in detection of 2-D partially visible objects in a complex scene is presented. Dominant local features (landmarks) which are rich enough in information to characterize the shape of an object are selected as the high curvature points along boundary of the object. The landmarks are extracted from both the model and the scene. The landmarks of the model are matched against those of the scene using a continuous Hopfield neural network in which feature matching is formulated as the optimization of a cost function. The sphericity of the triangular transformation, which is invariant with respect to translation, rotation, and scaling, is used as a similarity measure in the feature matching module. The matched landmarks resulting from the feature matching module are introduced to the decision making module which evaluates the overall goodness of the match and identifies the location of the object in the scene. >