Skip to content

4.0 End‐to‐end Optical Music Recognition

Jump to bottom
Geneviève Gates-Panneton edited this page Aug 20, 2024 · 5 revisions

After Pixel and the Interactive Classifier comes the last step: the end-to-end Optical Music Recognition, or e2e OMR, workflow! This is where the computer will analyse a folio image to produce its own transcription in MEI format. Before beginning, it's highly recommended to check your resource names and labels and make sure you're able to distinguish between them/easily identify what each file contains. Onwards!

For this step, you will need:

  • The picture of the given folio;
  • The three layer models produced by the Pixel job: Model 1 for glyphs, model 2 for staff lines, and the background model, which includes the text;
  • The training data produced by the Interactive Classifier;
  • A text file containing the text of all the chants on the given folio. This text file can be extracted from https://cantusdatabase.org/;
  • A csv file containing a spreadsheet assigning every glyph category its correct encoding in MEI format. This file can be produced using https://cress.simssa.ca/. An explanation of how the MEI format works can be found here.

The workflow is built like so:

NEW OMR jobs

For more information on the three branches of the workflow and how they connect to each other and work, please consult the introductory chapter on the OMR workflow.

Once the workflow is built, the resources can be uploaded. Resources are assigned like so:

NEW OMR ports

It's possible to process multiple folios in one workflow run by assigning multiple folio pictures to the original PNG job, as well as their corresponding text files to the Text Alignment job.

Important

Make sure when you do this that the pictures and text files are in the same order, so that they're matched to each other correctly.

Unlike the two previous steps, the e2e OMR workflow requires no human input at all. The given folio is read and transcribed using the training data and models produced with Pixel and the IC; all that's needed of the human is to connect the jobs, input the required resources, and click 'Run'!

This workflow run may take some time, especially if multiple images are being processed at once. Once it's complete, you'll be able to download the resulting MEI file (or files) from the 'Resources' tab.

Add more to this chapter when we start running OMR workflows again

MEI files can then be corrected using Neon. For more information on that, please visit the Neon wiki.

Clone this wiki locally