Abstract

Optical Music Recognition (OMR) is a branch ofartificial intelligence that aims at automatically recognizingand understanding the content of music scores in images.Several approaches and systems have been proposed that try tosolve this problem by using expert knowledge and specializedalgorithms that tend to fail at generalization to a broaderset of scores, imperfect image scans or data of differentformatting. In this paper we propose a new approach to solveOMR by investigating how humans read music scores and byimitating that behavior with machine learning. To demonstratethe power of this approach, we conduct two experimentsthat teach a machine to distinguish entire music sheets fromarbitrary content through frame-by-frame classification anddistinguishing between 32 classes of handwritten music symbolswhich can be a basis for object detection. Both tasks canbe performed at high rates of confidence (>98%) which iscomparable to the performance of humans on the same task.

Reference

Pacha, A., & Eidenberger, H. (2017). Towards Self-Learning Optical Music Recognition. In 2017 16th IEEE International Conference on Machine Learning and Applications (ICMLA). 16th IEEE International Conference on Machine Learning and Applications, Cancun, Non-EU. https://doi.org/10.1109/icmla.2017.00-60