Learning Spatial Grammars for Drawn Documents Using Genetic Algorithms

The problem of object recognition may be cast into a spatial grammar framework. The system comprises three novel elements: a spatial organisation of line features, an efficient two dimensional parsing engine, and a genetic algorithm learning routine that induces spatial grammars. Labelling the spatial organisation of feature pairs allows the terminal symbols of the spatial grammar to be defined, and constrains the search space of the feature parser. A genetic algorithm approach is then used to induce appropriate grammars using a supervised learning routine. Early results show that similar foreground and background features can be discriminated using this approach.