Hypothesis generation and correction in a hand-written word recognition system

A new method for word construction and hypothesis generation, implemented in a system for machine reading of USA Census '90 hand-written forms, is introduced. Several heuristic tests control the building of possible character images based on primitive image blobs. Confidence levels, returned by an isolated character classifier are combined with other contextual information to associate a degree of belief with possible words, facing the problem of skipping some artifacts, such as those often found in images produced by scanning microfilms. Inexact dictionary matching techniques are proposed to reduce the recognition error due to the incorrect detection of character boundaries. From the analysis of the NIST Special Database 12 we have observed that almost one third of the hypotheses are corrected by the inexact matching dictionary correction module, and 55% of these hypotheses are classified correctly.