Vectorial Signatures for Symbol Discrimination

In this paper, we present a method based on vectorial signatures, which aims at discriminating, by a fast technique, symbols represented within technical documents. The use of signatures on this kind of document has an obvious interest. Indeed, considering a raw vectorial description of the graphical layer of a technical document (e.g. a set of arcs and segments), signatures can be used to perform a pre-processing step before a "traditional" graphics recognition processing, or can be used to establish a classification that can be sufficient to feed a further indexation step. To compute vectorial signatures, we have based our approach on a method proposed by Etemadi et al., who study spatial relations between primitives to solve a vision problem. We consider five types of relations, invariant to transformations like rotation or scaling, between neighboring segments: parallelism with or without overlapping, collinearity, L junctions and V junctions. A quality factor is computed for each of the relations, computable with low requirements of power. The signature of all models of symbols that could be found in a given document are computed and matched against the signature of the document, in order to determine what symbols the document is likely to contain. The quality factor associated with each relation is used to prune relations whose quality factor is too low. We present finally the first tests obtained with this method, and we discuss the improvements we plan to do.