Symbolic indirect correlation classifier.

摘要

Symbolic Indirect Correlation (SIC) is a new, general method for non-parametric classification of unsegmented patterns that have local correlation with their labels. SIC uses two tiers of comparisons. At the first level, the feature-string representation of the unknown signal is compared to the feature-string representation of the reference signal of known words or phrases, and each class (word) in a lexicon of allowable classes is compared to the transcript of the reference set. At the second level, the properties of the feature-level matches of the unknown signal with the reference is compared with the properties of matches of each lexicon class with the transcript of the reference set. The unknown pattern is classified according to the best matching lexical class in the second comparison.One way of viewing the SIC problem is to find the correspondence, if one exists, between two bipartite graphs, one representing the matching of the two lexical strings and the other representing the matching of the two signal strings. For excessively noisy situations where graph matching becomes inefficient, it is possible to pose this as a maximum likelihood classification problem. We present here the mathematical formulation of the classification problem using both approaches and show the feasibility of the method for Optical Character Recognition (OCR) on simulations and online handwriting.