Graphical symbol recognition in architectural plans with an improved Ant-Tree based clustering algorithm

In this paper, an improved clustering algorithm based Ant-Tree is used for recognition of certain kind of architectural symbols with prior knowledge in engineering drawings. Symbols are segmented from an AutoCAD format drawing and a vector of invariants based on pseudo-Zernike moments is calculated to represent the graphical feature of a symbol. A normalization method is used to make these moments invariant of translation, rotation and scaling. Then the improved Ant-Tree algorithm is applied to cluster the symbols with regard to their features. The class of target symbols can thus be got easily with the guidance of some prior knowledge. For the proposed clustering algorithm, a new initialization method is presented with regard to the distribution of the data, and centroid approximation is also utilized to optimize the clustering process. Experiments show the effectiveness of our recognition approach proposed.