Extracting Words and Multi-Part Symbols in Graphics Rich Documents

We present an algorithm for grouping multipart symbols dashed lines and character strings for extraction from line drawings The image under goes a lossless raster to vector conversion creating as its vector representa tion an undirected graph a so called run graph Next the image elements of the run graph are extracted and classi ed probabilistically based upon their geometric features using a decision tree An area Voronoi tessella tion of the members of the sets is constructed from which a neighborhood graph is derived which is guaranteed to be minimal and complete The graph is then traversed to group the members of the various sets for ex traction and input to di erent recognition modules No a priori font or other domain speci c information is required for the grouping and no spe cial geometrical relationships among the elements are assumed Results are presented with example images taken from those used by our Swiss cadastral map understanding system