C. Flanagan, M.A. Rahman and E. McQuade
Synopsis
In this paper we investigate and analyse the behaviour of N-tuple pattern classifiers in noisy conditions. In our study we have found that the ability of an N-tuple classifier to classify patterns correctly is strongly dependent upon both the amount of noise in an image and the tuple size used during classification, with small tuple sizes performing markedly better than large ones. We have developed a mathematical model to account for this behaviour and have shown that this model is in good agreement with the experiments employing single train patterns. Based upon this mathematical model we have suggested a modification to the N-tuple classifier architecture which we call "oversampling", and have experimentally demonstrated its efficacy as a means of improving classification performance.
Key words:
Pattern Classification, Neural Networks, N-Tuple Techniques, Machine Vision