Improved image classification using morphing

Abstract

Principal component methods for classifying images have received broad attention and application. For objects with varying appearance, such as three-dimensional objects, increasing the number of object poses represented in the training set is the primary method for improving classification rate. In this paper we show how to improve the performance of this kind of an appearance-based image recognition system. The improvement is obtained by adding new views to the training set which have been generated from existing training data via a morphing algorithm. We show that adding morphed views to the training set increases recognition rate over the same data without morphed views.