In this article, we have tried to explore a new hybrid approach which well integrates the advantages of structural and statistical approaches and avoids their weaknesses. In the proposed approach, the graphic symbols are first segmented into high-level primitive like quadrilaterals. Then, a graph is built by utilizing these quadrilaterals as nodes and their spatial relationships as edges. Additional information like relative length of the quadrilaterals and their relative angles with neighbouring quadrilaterals are associated as attributes to the nodes and edges of the graph respectively. However, the observed graphs are subject to deformations due to noise and/or vectorial distortion (in case of hand-drawn images) hence differs somewhat from their ideal models by either missing or extra nodes and edges appearance. Therefore, we propose a method that computes a measure of similarity between two given graphs instead of looking for exact isomorphism. The approach is based on comparing feature vectors extracted from the graphs. The idea is to use features that can be quickly computed from a graph on the one hand, but are, on the other hand, effective in discriminating between the various graphs in the database. The nearest neighbour rule is used as a classifier due to its simplicity and good behaviour.
展开▼
