Abstract

Many proposed architectures for optical pattern-recognition use large space–bandwidth template matching. Experience gained in the last 10 years or so of research indicates that a large number of templates/filters are required in order for most applications to be practical. Traditionally, optical-correlation techniques have relied on templates/filters being either serially presented in time or spatially multiplexed in two dimensions. For many applications, the former approach is penalized by excessive time–overhead, whereas the latter approach requires an unrealistically large space–bandwidth product. In this paper, we describe the architecture and functioning of a massively parallel optical template matcher/correlator capable of calculating the inner products of 1,000 high-space–bandwidth pattern vectors simultaneously. Our template matcher/correlator consists of a large number of angle-multiplexed, Fourier-space volume holograms stored in a LiNbO₃:Fe crystal. Each hologram is stored by exposing the LiNbO₃ crystal to the interference pattern of one of the templates/filters to be stored and to a plane wave. Each plane wave has a slightly different (horizontal) angle of incidence for each hologram. All templates/filters, however, are incident at the same angle, Θ. When an input image illuminates the crystal at angle Θ, partial reconstructions of the stored plane waves, each proportional to the cross correlations of the input image with the stored templates/filters, emerge from the crystal. These plane waves are then focused by a spherical lens to form a one-dimensional array of pattern vector inner products and cross correlations (with one-dimensional shift invariance) in the orthogonal dimension. We have experimentally generated the covariance matrix for a 1000- template/filter device and have found good agreement with the corresponding, numerically generated matrix.

© 1990 Optical Society of America