c-jordi / pdf2data

A pdf segmentation and annotation tool for archival documents.

pdf2data : A pdf segmentation and annotation tool for archival documents.

🚧 IN DEVELOPMENT 🚧

From the Chair of Systems Design, ETHZ and the Swiss Data Science Center.

💡 Vision

Develop an open source and user-friendly tool for technical and non-technical users that performs page, block & textline segmentation and combines both manual & automatic annotation.

🎥 Preview

🔥 Features

• Structure your work into project and case studies.
• Upload your pdf files.
• Annotate the results of the segmentation algorithm using the interactive dashboard.
• Automate the training of a classification algorithm.
• Export your results for further analysis.

🚀 Quickstart

Development

Start the message broker:

docker-compose up

Start the backend:

source server/venv/bin/activate
make run

Start the worker:

source server/venv/bin/activate
make worker

Start the client:

cd client
yarn start

🧮 Data Composition

📚 Stack

• Node.js and React.js deliver the interactive dashboard.
• Tornado runs the data backend.
• Celery with a RabbitMQ backend operates the execution of asynchronous tasks.

pdf2data's Issues

Add all possible font types during feature extraction

So far, only a handful of font types (Calibri, Times New Roman, etc.) are considered for the extraction of features. However, we may encounter many different fonts. We need to implement the following:

• During feature extraction, identify all possible font types
• After processing all documents, run another step, over all the fonts encountered, to extract features (font occurrence) for all the types.

This feature will be part of the feature extraction process, contained in server/application/feature.

Enrich Pages table in DB

We need to create one entry per page in the Table DB. The keys to fill in are quite straightforward, and the only question that arises is, when to do it??

• I do believe the best would be to just do it after the corresponding XML file is extracted. Hence, from the XML we read the number of pages, and we just create one entry per page.
• But we could also do it after the extraction of features.

Since this is just a matter of commit this to the DB, could you decide @c-jordi when is the best moment and implement it?
Thanks!

Add even more hand crafted features

Despite we will add more complex models to extract features, perhaps add the following extra, to each of the tree levels

• Pages: percent of page covered, normalized page number
• Textbox: TBD
• Lines: TBD

Issue with Textline feature extraction

Running the feature extraction with textline leads to a ValueError.

ValueError: Found array with 0 sample(s) (shape=(0, 1)) while a minimum of 2 is required by AgglomerativeClustering.