Novel neural network models for computing homothetic invariances: an image algebra notation

Abstract

In this paper we propose a theoretical approach toinvariant perception. Invariant perception is an importantaspect in both natural and artificial perception systems, and itremains an important unsolved problem in heuristically basedpattern recognition. Our approach is based on a general theoryof neural networks and studies of invariant perception by thecortex. The neural structures that we propose uphold both thearchitecture and functionality of the cortex as currentlyunderstood. The formulation of the proposed neural structuresis in the language of image algebra, a mathematical environmentfor expressing image processing algorithms. Thus, an additionalbenefit of our study is the implication that image algebraprovides an excellent environment for expressing and developingartificial perception systems. The focus of our study is oninvariances that are expressible in terms of affinetransformations, specifically, homothetic transformations. Ourdiscussion will include both one-dimensional andtwo-dimensional signal patterns. The main contribution of thispaper is the formulation of several novel morphological neuralnetworks that compute homothetic auditory and visualinvariances. With respect to the latter, we employ the theoryand trends of currently popular artificial vision systems.