Abstract

We show that image restoration used as a preprocessing tool improves rotation invariant pattern recognition. The effects of signal-dependent random noise on rotation invariant pattern recognition were investigated. Computer simulations were performed on 10,000 noisy realizations of each member of a set of binary and grey level images using circular harmonic, phase-only circular harmonic, and binary phase-only circular harmonic filters. A significant number of false detections have been observed in the correlation plane, thus establishing the filters as less robust with respect to noise than the corresponding non rotation invariant matched filters.¹ The noisy images giving rise to false detections were restored by locally adaptive filters. We show that, as a result, false detections were eliminated.

© 1991 Optical Society of America