A Deep Approach for Handwritten Musical Symbols Recognition

• Roberto M. Pinheiro Pereira UFMA
• Caio Eduardo Falcao Matos UFMA
• Geraldo Braz Junior UFMA
• João Dallyson Sousa de Almeida UFMA
• Anselmo Cardoso de Paiva UFMA

Resumo

Preserving the world musical heritage comes down to digitalizing and provision of music works to further query on the acquired data. However, to do the processing it is necessary an Optical Music Recognition (OMR) system capable of decoding the original manuscripts into a machine-readable data. Developing a precise and robust OMR system for handwritten musical scores is still an open issue. A fundamental step of improve such task is to recognise musical notes. Hence, trying to provide ways to produce a truly robust OMR system, we present in this paper a new methodology applying deep learning techniques to recognise musical notes in digitalised handwritten musical scores. The proposed methodology has been tested on a ground truth dataset of music scores reaching a minimum error rate of 3.99%, 96.46% of precision and 96.56% of recall on the HOMUS dataset.