Abstract: Existing methods of recognition and understanding of architectural drawings suffer from the mitations of sensitivity to noise and bad generality, which makes precise analysis and econstruction much more difficult. Analyzing the semantic content of architecture and simulating the way human read architectural drawings, an efficient self-incremental axis-net-based hierarchical recognition (SINEHIR) model is proposed. SINEHIR recognizes each class of architectural components sequentially, driven by recognized components in the drawing which is incrementally simplified by eliminating the interference of previous recognition. Also presented are feature-based symbol recognition method, symbol-based axes net recognition method, section-tracking-based node recognition method, semantic-relation-based segment recognition method, geometry-based combined component recognition method, and inheritance-based data transition method of SINEHIR.