Abstract
The new non-parametric approach to unsegmented text recognition builds two bipartite graphs that result from the feature-level and lexical comparisons of the same word against a reference string which need not include the query word. The lexical graph preserves the relative order of edges in the feature graph corresponding to correctly recognized features. This observation leads to a subgraph-matching formulation of the recognition problem. An initial implementation proves the robustness of the approach for up-to 20% noise introduced in the feature-level graph.