Structured reconstruction of CAD floor plans based on topological constraints and adaptive parameter optimization

Abstract

Structured reconstruction of CAD floor plans is a core component of automatic Building Information Modeling (BIM) generation, indoor spatial analysis, and the digitalization of architectural heritage. To address the key limitations of traditional image processing methods-namely high parameter sensitivity and poor robustness-as well as the reliance of deep learning approaches on large-scale, professionally annotated datasets, this paper proposes a multi-stage CAD floor plan structured reconstruction framework that integrates topological constraints and adaptive parameter optimization. The proposed framework sequentially performs image preprocessing, wall extraction and enhancement, door and window candidate detection with joint geometric-topological filtering, and room region closure restoration. A comprehensive evaluation function consisting of six topological metrics is designed, and an iterative rule-driven optimization mechanism is employed to automatically search for the optimal reconstruction result. Experiments conducted on a dataset of 20 residential CAD floor plans with diverse styles demonstrate that the proposed method improves wall connectivity by 32.7%, achieves an 87.2% wall-door/window attachment rate, and reaches 100% room closure rate. Overall performance outperforms both traditional fixed-parameter methods and mainstream unsupervised segmentation approaches.