ABSTRACT
Ever since the first use of the optical correlator for implementing matched spatial filtering [1], many studies have been made to develop new algorithms that increase the performances of the correlation filters. Correlation filters are generally computed on the basis of reference images containing the objects to be recognized or rejected. The set of reference images is chosen in order to cover the different occurrences of the objects in the best way. However, in practical applications, input images may differ strongly from their reference model. In this case, the maximum detected intensity in the correlation plane decreases, which may lead to false detections. In order to have an efficient recognition system robust to input distortions, one must be able to realize filters insensitive to pattern distortions, illumination problems, and additive and substitutive noise appearing in the input images.
