基于深度学习的公共建筑像素施工图空间识别

Deep learning based pixel-level public architectural floor plan space recognition

像素级图纸空间识别在自动审图、图纸翻模应用中起到关键作用。已有识别方案主要针对户型图,基于语义分割直接识别空间。公共建筑施工图有更多干扰线和图纸元素、更高分辨率及更多空间类型,高分辨率使模型难以获取图纸的全局信息,空间类型的多样性导致无法确定空间类型的范围,已有空间识别方案难以直接迁移。为了识别公共建筑施工图中的空间,标注公共建筑施工图数据集,包含20张标注墙体像素、100张标注构件包围盒的公共建筑施工图,提出基于深度学习的空间边界识别方案及中心线提取和关键线均方误差损失函数,以提高空间边界中墙体的识别准确率,通过空间围合识别空间,提出空间边界平滑算法,在保持空间形状的同时减少空间边界点数量。实验结果表明,该方法突破了分辨率和空间类型限制,取得了较好的空间识别结果,为识别公共建筑施工图的空间提出了解决方案。相较于已有方法,该方法在保证墙体精确率的情况下获得了更高的召回率。

Pixel-level floor plan space recognition plays an important role in applications such as floor plan review and model reconstruction from drawings. Targeting at housing floor plans, the existing methods recognize spaces directly based on semantic segmentation. Public architectural floor plans feature more noising lines and elements,higher resolution, and more space varieties. Higher resolution makes it hard to acquire global information in a floor plan, while the variety of spaces makes it impossible to gain the clear range of room types, both features rendering the existing space recognition approaches unpractical. To recognize spaces in public architectural floor plans, a dataset named Public Architectural Floor Plan Dataset was proposed, including 20 floor plans labeled with walls at the pixel level and 100 floor plans labeled with elements at the bounding box level. A deep learning-based space boundary recognition approach was proposed. This approach could enhance the accuracy in recognizing walls, with the proposed center line extraction and key line minimum square error loss function, and could recognize spaces by enclosing space. A space contour optimization algorithm was proposed, which in experiments could reduce the number of contour points and reserve the shape of spaces. Experimental results show that this method breaks through the limitation of resolution and room type range, attains satisfying space recognition performance, and presents a solution to recognizing spaces of public architectural floor plans. Compared with existing methods, the proposed method reaches a higher recall ratio while the precision score is guaranteed.