| Summary: | In statistical pattern recognition, issues on large learning samples to be used with the well-known, k-Nearest Neighbour (k-NN) classifier, have been investigated by many researchers. These prolems are due to the the large computational burden and memory space required to acomodate these samples. various editing dand condensing techniques have been implemented to reduce the number of learning samples. In some recognition applications, it is difficult to obtain large learning samples; thus in many research, the k-NN rule is modified to suit the requirements of the application. We will study a new approach to avoid the use of large learning samples by finding a good distance measure. We employ teh Nearest Prototype (NP) classifier, a prototype based decision rule that works on the same principles as the k-NN rule for k=1. We investigate various types of distance measures via different matching techniques based on the image representations. Performance of the NP Classifier iis indirectly related to the ggodness of the distance measure. A new method of evaluating the performance is also implemented in this thesis. From our experiments, a good distance measure based on pixel matching is found and its performance for 2-D symbol recognition is very promising.
|
|---|