Constrained composite filter for intraclass distortion invariant object recognition

A composite filter for a reliable recognition of a target and its intraclass distortions is presented. The filter is synthesized with the help of two data sets, a conventional training set and a proposed optimization set used to represent statistically the intraclass distortions. The proposed filter optimizes the mean squared error between the filter response to the intraclass distortions, and a prespecified correlation response over the training set. The resultant filter is able to satisfy the equal output correlation-peak constraints of conventional composite filters and at the same time to produce a statistically optimized response to the intraclass distortions. Computer simulations obtained with the proposed method in geometrically distorted and noisy scenes are discussed and compared with those obtained with existing composite filters in terms to tolerance to intraclass distortions and to the signal to noise ratio.