Cascaded linear shift invariant processing to improve discrimination and noise tolerance in optical pattern recognition

In this paper we report a study of optical pattern recognition using a cascade of linear shift invariant processing modules (correlators) each augmented with a thresholding layer. This configuration can be considered as a special class of multilayer feed-forward neural network. In contrast with more generalised multi-layer networks, the approach is easily implemented in practice using optical techniques and consequently well suited to the analysis of large images. The concept of cascaded linear shift invariant processing is introduced within the context of network analysis. It is shown that the system is equivalent to a multi-layer network which is constrained to have a shift invariant output. The system has been modelled using a modified back propagation algorithm with optimisation using simulated annealing techniques. The performance of the system has been compared to that of single layer correlators using a range of synthetic filters taken from the published literature. In particular we show that the noise tolerance of the cascaded system is increased relative to that of the minimum variance synthetic discriminant function (MVSDF). In addition we show that discrimination is enhanced considerably with respect to minimum average correlation energy (MACE) filters for the case of similar input images.