Deep Learning Applications in Architectural Floor Plan Analysis

Deep learning has transformed the interpretation of architectural floor plans by automating the extraction, classification and reconstruction of spatial information. Convolutional neural networks have been employed to segment raster floor-plan images into structural elements such as walls, doors and windows. Generative adversarial networks have further enabled the translation of pixel-based drawings into vector-formed primitives, facilitating interoperability with building information modelling frameworks. Developments in object detection and graph-based topology reconstruction consolidate discrete symbols and annotations into coherent spatial graphs. These advances address longstanding challenges in dataset scarcity, annotation complexity and class imbalance. Practical applications span heritage conservation, rapid prototyping of building layouts, emergency planning and integration into digital twins. The growing body of work signals a shift towards end-to-end pipelines that combine segmentation, symbol recognition and semantic parsing to deliver high-fidelity, machine-readable floor-plan representations.

Research from Nature Portfolio

No recent Nature Portfolio content available.

Research from all publishers

A recent comprehensive review of deep learning methods for digitisation of complex documents and engineering diagrams surveys symbol recognition, text extraction and connectivity detection across diverse sectors including architecture. It highlights the evolution from hand-crafted features to fully learned approaches, critiques the paucity of curated floor-plan datasets, and sets out a research agenda emphasising robust evaluation protocols and balanced class distributions.

An indoor reconstruction framework based on convolutional neural networks demonstrates accurate detection of wall and door pixels in floor-plan images, coupled with centreline and corner detection algorithms to convert segmented rasters into vector data. By preserving wall thickness and exporting to standards such as CityGML and IndoorGML, the approach enables seamless integration into three-dimensional spatial information systems, supporting urban modelling and facility management.

A vectorisation pipeline for two-dimensional floor plans employs an edge-extracting generative adversarial network to project raw floor-plan images into primitive feature maps. The introduction of a self-supervising loss term ensures fidelity of extracted structural lines. Downstream inspection modules based on subspace connectivity graphs and graph neural networks verify topological completeness and consistency with textual annotations, delivering high-speed vectorisation suitable for real-time design workflows.

Deep Learning Applications in Architectural Floor Plan Analysis publication trend

The graph below shows the total number of articles in deep learning applications in architectural floor plan analysis across all publications each year (not limited to Nature Index journals).

Technical terms

Convolutional Neural Network (CNN): A class of deep networks that apply convolutional filters to extract hierarchical features from images.

Generative Adversarial Network (GAN): A framework of two competing networks, one generating data and the other discriminating, used here to translate images into vector primitives.

Segmentation: The process of partitioning an image into regions corresponding to distinct architectural elements.

Vectorisation: Conversion of raster images into geometric primitives such as lines and polygons.

IndoorGML/CityGML: Open standards for encoding semantic and geometric information of indoor and urban spaces in vector form.

Topological Reconstruction: Assembly of detected primitives into a coherent graph that reflects spatial adjacency and connectivity.

References

1. A review of deep learning methods for digitisation of complex documents and engineering diagrams. Artificial Intelligence Review (2024).
2. Indoor Reconstruction from Floorplan Images with a Deep Learning Approach. ISPRS International Journal of Geo-Information (2020).
3. Vectorization of Floor Plans Based on EdgeGAN. Information (2021).