Efficient notation assembly in optical music recognition

Abstract

Optical Music Recognition (OMR) is the field of research that studies how to computationally read music notation from written documents. Thanks to recent advances in computer vision and deep learning, there are successful approaches that can locate the music-notation elements from a given music score image. Once detected, these elements must be related to each other to reconstruct the musical notation itself, in the so-called notation assembly stage. However, despite its relevance in the eventual success of the OMR, this stage has been barely addressed in the literature. This work presents a set of neural approaches to perform this assembly stage. Taking into account the number of possible syntactic relationships in a music score, we give special importance to the efficiency of the process in order to obtain useful models in practice. Our experiments, using the MUSCIMA++ handwritten sheet music dataset, show that the considered approaches are capable of outperforming the existing state of the art in terms of efficiency with limited (or no) performance degradation. We believe that the conclusions of this work provide novel insights into the notation assembly step, while indicating clues on how to approach the previous stages of the OMR and improve the overall performance.