An experimental comparison of one-class classification methods
Dick de Ridder, David M.J. Tax, Robert P.W. Duin
pattern recognition, one-class problems, neural networks, image segmentation, one class classification
This paper discusses several methods to perform one-class classification, i.e. classification in problems where only objects of one class are of real interest as opposed to all other possible objects. We compare a number of unsupervised methods from classical pattern recognition to a number of supervised neural classifiers. We will also introduce a new approach, which is a local combination of two traditional methods. As an experimental dataset we use samples taken
from scanned newspaper images. Using results from our experiments, the relative advantages and disadvantages of the different methods will be discussed.