MSc project investigating multi-modal fusion approaches to combining textual and visual features for multi-page document classification of documents within the North Sea Transition Authority (OGA) National Data Repository (NDR) using deep multimodal fusion convolutional long short term memory (C-LSTM) neural networks. This readme gives a brief overview of the project and the code in this repository, see the accompanying report for the full details. Note this is experimental code for my masters project, I would advise against running any of it in production.
All Python dependencies for the project can be installed by building a conda environment from the environment.yml file.
The data source used in this project was compiled from a corpus of raw documents uploaded by oil companies to the National Data Repository (NDR), a data repository for UK petroleum exploration and production data maintained by the North Sea Transition Authority (OGA). The document corpus used consists of a sample of 6,541 documents, these are mostly PDF, Microsoft Office, text and image type files.
The documents in this corpus are split into 6 classes:
- geol_geow - Geological end of well reports.
- geo_sed - Geological sedimentary reports.
- gphys_gen - General geophysical reports.
- log_sum - Well log summaries.
- pre-site - Pre-site reports.
- vsp_file - Vertical seismic profiles.