Improved minimum distance classification with Gaussian outlier detection for industrial inspection

Abstract

A pattern recognition system used for industrial inspection has to be highly reliable and fast. The reliability is essential for reducing the cost caused by incorrect decisions, while speed is necessary for real-time operation. We address the problem of inspecting optical media like compact disks and digital versatile disks. As the disks are checked during production and the output of the production line has to be sufficiently high, the time available for the whole examination is very short, ie, about 1 sec per disk. In such real-time applications, the well-known minimum distance algorithm is often used as classifier. However, its main drawback is the unreliability when the training data are not well clustered in feature-space. Here we describe a method for off-line outlier detection, which cleans the training data set and yields substantially better classification results. It works on a statistical test basis. In addition, two improved versions of the minimum distance classifier, which both yield higher rates of correct classification with practically no speed-loss are presented. To evaluate the results, we compare them to the results obtained using a standard minimum distance classifier, a k-nearest neighbor classifier, and a fuzzy k-nearest neighbor classifier.