Search for a command to run...
With the rapid advancement of building informatization and smart cities, floor plan recognition has become a critical bridge between two-dimensional drawings and three-dimensional visualization models. However, floor plan parsing still faces challenges such as insufficient data diversity, dense non-structural interference, and difficulties in fine-grained component recognition, making traditional image-processing methods inadequate for high-precision analysis. To address these challenges, a scene-adaptive intelligent floor plan parsing approach is proposed, integrating object detection and semantic segmentation to achieve end-to-end recognition and geometric reconstruction of walls, doors, windows, and interior spaces. On the data side, a multi-source dataset is constructed, covering public drawings, industrial CAD blueprints, and real estate floor plans, with an optimized preprocessing pipeline involving slidingwindow partitioning and primitive cleaning. On the model side, an integrated interference removal and structural segmentation method is designed, in which YOLOv8 is employed to eliminate non-structural graphic elements, and Mask2Former is adopted for high-precision segmentation of walls and room regions. Furthermore, for component recognition and post-processing, a YOLOE-based detector combined with visual prompting achieves accurate fine-grained component detection, while a geometrydriven post-processing module enhances structural closure and usability. Extensive experiments demonstrate that the proposed method achieves up to 97.6 % wall segmentation accuracy, 98.3 % room mIoU, and 99.1 % component detection mAP, significantly outperforming mainstream baselines in structural connectivity, component recognition, and robustness across diverse architectural scenarios.