Introduction

Although image features can be characterised to some extent by intrinsic attributes such as local image gradients and curvatures the context of the surrounding shape geometry can provide the basis for a much more powerful descriptor. By selecting appropriate parameters and storing these measurements in the form of a frequency histogram, a concise shape descriptor can be produced which promotes robust feature classification. This histogram is referred to as a pairwise geometric histogram because it records geometric measures made between pairs of image features [1], [2].

Built into the PGH paradigm is a set of parameters defining the histogram type, accuracy and quantization of the axes of the histograms. Presently some of these parameters are decided using rules of thumb and trial-and-error. This paper describes the application of a MOGA to determine these parameters in a principled manner. The significance of this work is that it enables the process of pairwise object recognition to be automatically extended for use on large generic model databases.