Rough sets in hybrid methods for pattern recognition

Many papers describe hybrid methods used for pattern recognition. Such systems consist of an optical part, performing fast signal preprocessing, and a computer, analyzing preprocessed data. Here we present. the method which uses for feature extraction the ring-wedge detectors (RWD) or computer generated holograms (CGH) serving as RWD. Features obtained in this way are shift, rotation, and scale invariant, but papers suggest that they can be still subject for further optimization. This article presents an original method for optimizing feature extraction abilities of CGH. This method uses rough set theory (RST) to measure the amount of essential information for the classification, preserved in feature vector. As there is no gradient direction information in factors defined by RST, we use for a space search a stochastic evolutionary approach. Finally, we use RST to determine decision rules for the feature vector classification. The whole method is illustrated by a system recognizing the speckle pattern images obtained as a result of interference of light going through a quasi-monomode optical fiber. As the conditions of interference differ when some kind of distortion of the optical fiber is produced, such a system can be used as a sensor of the pressure causing this distortion. (C) 2000 John Wiley & Sons, Inc.