A neural approach for full-page optical music recognition of mensural documents

The digitization of the content within musical manuscripts allows the possibility of preserving, disseminating, and exploiting that cultural heritage. The automation of this process has been object of study for a long time in the field of Optical Music Recognition (OMR), with a wide variety of proposed solutions. Currently, there is a tendency to use machine learning strategies based on neural networks because of their high performance and flexibility to adapt to different scenarios by changing only the training data. However, most of the recent literature addresses only specific parts of the traditional OMR workflow such as music object detection or symbol classification. In this paper, we progress one step further by proposing a full-page OMR system for Mensural notation scores that consists of simply two processes, which are enough to extract the symbolic music information from a full page. More precisely, our pipeline uses Selectional Auto-Encoders to extract single staff regions, combined with end-to-end staff-level recognition based on Convolutional Recurrent Neural Networks for retrieving the music notation. The results confirm the adequacy of our method, reporting a successful behavior on two Mensural collections (Capitan and Seils datasets) with a straightforward implementation