VisArchPy

Data pipelines for extraction, transformation and visualization of architectural visuals in Python. It extracts images embedded in PDF files, collects relevant metadata, and extracts visual features using the DinoV2 model. We ambition to make of this package Ai-powered tool with features for recorgnizing different types architectural visuals (types of buildings, structures, etc.). The package is still in development and we are working on adding more features and improving the existing ones. If you have any suggestions or questions, please open an issue in our GitHub repository.

Main Features

Extraction pipelines

• Layout: pipeline for extracting metadata and visuals (images) from PDF files using a layout analysis. Layout analysis recursively checks elements in the PDF file and sorts them into images, text, and other elements.
• OCR: pipeline for extracting metadata and visuals from PDF files using OCR analysis. OCR analysis extracts images from PDF files using Tesseract OCR.
• LayoutOCR: pipeline for extracting metadata and visuals from PDF files that combines layout and OCR analysis.

Metadata Extraction

• Extraction of medatdata of extracted images (document page, image size)
• Extraction of captions of images based on proximity to images and text-analysis using keywords.

Transformation utilities

• Dino: pipeline for transforming images into visual features using the self-supervised learning in DinoV2.

Visualization utilities

• Viz: an utility to create a bounding box plot. This plot provides an overview of the shapes and sizes of images in a data set.

  

Dependencies

• Python 3.10 or 3.11
• Tesseract v4.0 or recent
• PyTorch v2.1 or recent

Installion

After installing the dependencies, install VisArchPy using pip.

shell pip install visarchpy

Installing from source

1. Clone the repository. shell git clone https://github.com/AiDAPT-A/VisArchPy.git
2. Go to the root of the repository. shell cd VisArchPy/

3. Install the package using pip.

   shell pip install .

Developers who intend to modify the sourcecode can install additional dependencies for test and documentation as follows.

1. Go to the root directory visarchpy/

2. Run:

shell pip install -e .[dev]

Usage

VisArchPy provides a command line interface to access its functionality. If you want to VisArchPy as a Python package consult the documentation.

1. To access the CLI:

shell visarch -h

1. To access a particular pipeline:

shell visarch [PIPELINE] [SUBCOMMAND]

For example, to run the layout pipeline using a single PDF file, do the following:

shell visarch layout from-file <path-to-pdf-file> <path-output-directory>

Use visarch [PIPELINE] [SUBCOMMAND] -h for help.

Results

Results from the data extraction pipelines (Layout, OCR, LayoutOCR) are save to the output directory. Results are organized as following:

shell 00000/ # results directory ├── pdf-001 # directory where images are saved to. One per PDF file ├── 00000-metadata.csv # extracted metadata as CSV ├── 00000-metadata.json # extracted metadata as JSON ├── 00000-settings.json # settings used by pipeline └── 00000.log # log file

Settings

The pipeline's settings determine how visual extraction from PDF files is performed. Settings must be passed as a JSON file on the CLI. Settings may must include all items listed below. The values showed belowed are the defaults.

\ When no seetings are passed to a pipeline, the defaults are used. To print the default seetting to the terminal use:

shell visarch [PIPELINE] settings

Citation

Please cite this software using as follows:

Garcia Alvarez, M. G., Khademi, S., & Pohl, D. (2023). VisArchPy [Computer software]. https://github.com/AiDAPT-A/VisArchPy

Acknowlegdements

• VisArchPy was develped thanks to the support provided by the Digital Competence Centre, Delft University of Technology.
• Reseach Data Services, Delft University of Technology, The Netherlands.