Training with positive and negative data samples: effects on a classifier for hand-drawn geometric shapes

Abstract

This thesis examines the use of positive and negative training data. on a nearestneighbour classifier for hand-drawn geometnc shapes. to improve reliability. A reliiible classifier must feature the ability to reject miss-segrnented and unknown shapes (ie.negative symbols). A recognition system's performance hinges on the performance and reliability of its classifier. In diagram recojnition, where the crowding of line-segments and text often causes segmentation errors. a method for rejecting negative symbols is necessary to improve the performance of the system. This allows feedback from the classifier and contextual information to be used to re-evaluate segmentation hypotheses.We explore the issues involved in the development of a reliable classifier by training it on both positive and negative data. and we discuss the trade-off between reliability and correctness.