[Back]

A. Pacha, J. Calvo-Zaragoza, J. Hajič:
"Learning Notation Graph Construction for Full-Pipeline Optical Music Recognition";
Talk: 20th International Society for Music Information Retrieval Conference, Delft, The Netherlands; 2019-11-04 - 2019-11-08; in: "20th International Society for Music Information Retrieval Conference", (2019), 75 - 82.

Optical Music Recognition (OMR) promises great benefits to Music Information Retrieval by reducing the costs of making sheet music available in a symbolic format. Recent advances in deep learning have turned typical OMR obstacles into clearly solvable problems, especially the stages that visually process the input image, such as staff line removal or detection of music-notation objects. However, merely detecting objects is not enough for retrieving the actual content, as music notation is a configurational writing system where the semantic of a primitive is defined by its relationship to other primitives. Thus, OMR systems must employ a notation assembly stage to infer such relationships among the detected objects. So far, this stage has been addressed by devising a set of predefined rules or grammars, which hardly generalize well. In this work, we formulate the notation assembly stage from a set of detected primitives as a machine learning problem. Our notation assembly is modeled as a graph that stores syntactic relationships among primitives, which allows us to capture the configuration of symbols in a music-notation document. Our results over the handwritten sheet music corpus MUSCIMA++ show 95.2% precision, 96.0% recall, and an F-score of 95.6% in establishing the correct syntactic relationships. When inferring relationships on data from a music object detector, the model achieves 93.2% precision, 91.5% recall and an F-score of 92.3%.

Optical Music Recognition, Deep Learning, Notation Graph Construction, Machine Learning

https://publik.tuwien.ac.at/files/publik_282965.pdf