Abstract
Detecting music notation symbols is the most immediate unsolved subproblem in Optical Music Recognition for musical manuscripts. We show that a U-Net architecture for semantic segmentation combined with a trivial detector already establishes a high baseline for this task, and we propose tricks that further improve detection performance: training against convex hulls of symbol masks, and multichannel output models that enable feature sharing for semantically related symbols.
The latter is helpful especially for clefs, which have severe impacts on the overall OMR result. We then integrate the networks into an OMR pipeline by applying a subsequent notation assembly stage, establishing a new baseline result for pitch inference in handwritten music at an f-score of 0.81.
Given the automatically inferred pitches we run retrieval experiments on handwritten scores, providing first empirical evidence that utilizing the powerful image processing models brings content-based search in large musical manuscript arc