Object Recognition by Selective Focusing Using a Moore-Penrose Associative Memory

Abstract

This paper describes a rotation-, scale-, and distortion-invariant method of 2D object recognition based on the Moore-Penrose Distributed Associative Memory (DAM). Invariance to rotation and scale is achieved using conformal mappings. For distortion invariance the memory is adpated to the key stimulus before recall which leads to improved accuracy for both recall and classification. Based on known relationships between DAMs and regression analysis, a method is proposed to improve the selectivity of the association weights in an iterative way by discarding insignificant response vectors. Such selectivity allows the system to focus on the significant associations and to reduce cross-talk effects. The same formalism which allows to compute the significance of the association weights also provides for a reject option. Experimental results prove the feasibility and benefits of the new recognition method.