This docker image builts an API which categorizes architectural drawings through Convolutional Neural Networks. The goal was to develop a framework to automatically create relevant metadata of drawings for the purpose of long-term archiving. The API takes a user input parsed via the HTTP interface in server.py, runs predictions on n number of images and stores the predictions in either a JSON-string or a csv-file.
It is an ongoing research from the FID BAUdigital, conducted at the Universitäts- und Landesbibliothek Darmstadt, at the TU Darmstadt.
PredictDrawingCategory
This script predicts if an architectural drawing either belongs to the category of a floor plan, a section or an elevation through a Multi-Class Classification Model. The accepted drawing types are therefore: floor plan, elevation and section.
To train the network on drawing categories, the following network architecture was used:
PredictDesignPattern
This script estimates the presence of certain architectural design patterns in floor plan drawings, built on top of my DeepPattern repository. Categories/Patterns include the following shapes: Rectangle, Circle, Composite-rectangular, Organic, Longitudinal, Polygonal. It furthermore includes the following interior Patterns: Atrium, Column Grid and Staircase. Accepted drawing types are thus only floor plans.
Usage
To run the API:
git clonedocker build -t python3 -f build/package/Dockerfiledocker -p 8080:8080 run python3http://localhost:8080/pattern or http://localhost:8080/category
For questions, email me at: paul.arch@web.de