Abstract

A critical task of machine vision is to identify and locate known reference images in the field of view. Although the simple matched filter is the best for detecting an exactly known signal in additive noise, its performance degrades rapidly as the reference image undergoes distortions such as rotation and scale change. One class of methods suggested for combatting this problem is known as synthetic discriminant functions (SDF)[1]. In this, an image is synthesized such that it yields constant cross-correlation values with a set of training images. If this training set is sufficiently representative of the type of distortions to be encountered, we can expect that this approach yields constant cross-correlation values when tested with non-training images. Recently, much attention has been directed towards increasing the repertoire of the SDF techniques beyond the basic scheme in which the synthesized image is assumed a priori to be a linear combination of the training images. In this paper, we propose a new method in which the cross-correlation outputs are allowed to be complex and which minimizes the output variance. This new technique offers significant reductions in the noise sensitivity of the conventional SDFs.

© 1987 Optical Society of America