Spatial Topology Distance for Handprinted Character Recognition

Abstract

In this work we introduce new distance measurements for calculating spatial topology dissimilarity and apply them in recognition of handprinted characters. In many applications (e.g. recognition of deformed images, handprinted characters, signatures, biomedical and geophysical patterns), it is of interest to analyze and recognize phenomena occuring at varying deformations. The recently introduced neural network models provide a learning-by-training and collective recognition capability that offer the possibility of such analysis and recognition. At present, however, there is no corresponding effective framework to support the development of effective neural network for rigorous real applications, such as handprinted character recognition for large set of characters. In this paper we provide such a framework. The devision of analog topology preserving map leads naturally to order the deformed patterns to their standard forms. Vector fields can be obtained from such ordering. The cumulation of each individual vector in the vector field is then properly defined and used as the perception energy. Several empirical perception energies are introduced for recognition of handprinted characters. An automatic recognition of mail address system which contains isolated Chinese characters, ten digits, and all English alphabets is developed based on such framework. A rigorous test is performed on the ten handprinted digits using their ten standard forms as reference database. More than eighty percent of the test database which contains 5000 different deformed handprinted digits can be correctly recognized. To our knowledge, this reference database is the smallest database for the same performance.