6x Ideas List

6x Ideas List

It is no surprise we are at a pivotal moment in the evolution of Optical Music Recognition.

For music symbols recognition, we are moving :

• from carefully isolated glyphs submitted to a shallow network trained on selected samples,
• to (sub/full)-images submitted to a deep network trained on a huge number of score images.

As detailed in 6.0 Prototype Presentation, this translates in Audiveris as:

• Release 5.1 just published in December 2018, with new UI tools to correct OMR errors, but still based on the old glyph classifier with its segmentation problems.
• 6.0 prototype documented at about the same moment, integrating a full-page detector/classifier and able to demo a patch classifier.

To move beyond this proof of concept, we have identified two main tasks that could be conducted in sequence or in parallel. One is the training of the new classifiers (page and patch), the other is their integration into Audiveris OMR. These are still ideas, to be discussed and refined.

Training task

Content

• Analyze Patch classifier behavior on notehead symbols.
• Augment current training data set (Lilypond-based synthetic scores, MUSCIMA++ hand-written scores) with MuseScore synthetic scores and Audiveris real printed scores.
• Provide trained models for Page and for Patch classifiers.

Technologies

• Python 3
• Tensorflow
• ResNet101

Integration task

Content

• Integrate Patch classifier in OMR engine to complement/correct Page results.
• Integrate Patch UI board to allow direct manual symbol assignment.

Technologies

• Java
• DeepLearning4J