Classification of Multi-Edge Shapes Using An Autoregressive Model and the Karhunen-Loeve Expansion

摘要

In this thesis a pattern recognition system capable of classifying two dimensional shapes with multiple edges was developed. The problem of multiple edge classification was treated as an extension of the single edge problem. For each edge, a feature vector was formed from the parameters of an autoregressive model of a time series representing the shape of the edge. The dimensional of these feature vectors was further reduced by the use of a transformation based on the Karhuene-Loeve expansion. A minimum distance classification rule was used to classify an input transformed feature vector according to the nearest class mean in the transformed feature space. Two boundary sampling methods as well as two versions of the Karhuene-Loeve transformation were investigated. An illustrative numerical example and the description of the system tests are provided. Using an equal angle boundary sampling technique and the pre-whitened Karhuene-Loeve transformation, an industrial shapes test showed 100% correct classification results with an average classification time of 1.27 seconds. The complete Fortran listings of the routines written for this system are included in the Appendix at the back of this work. (Author)