Vector Matching Methods for Large-Scale Pattern Recognition Tasks

Abstract

Pattern recognition tasks require an unknown object to be classified as one of a large number of patterns. The recognition framework is usually set up in a learning process in which a much larger number of training samples are handled. This article presents a series of data reduction methods that prove to be very effective for such tasks. First, an algorithm is introduced for constructing standard patterns, or templates, out of a training database. This algorithm ensures that each training sample falls in the attraction domain of a template of the same type, and the algorithm attempts to maintain a small number of templates. Due to the large quantity of templates thus derived, it is necessary to extract a subset for each unknown object. Thus, the next method is to build the storage devices, called template trees, for the purpose of fast candidate retrieval. A further candidate reduction can be made by a subvector matching technique, whose training method is also provided in this article. In many applications, one also faces the situation that certain types of patterns can easily be confused with each other. Thus, the third method presented in this article is to provide a tool for capturing those confusing patters in the training process and also to provide ways for their disambiguation.