Testing YOLOv8’s efficacy as a pitch and duration detector across digitally written monophonic music scores

Music presents itself as an important cultural landmark throughout human history that is largely recorded in physical sheets of paper that are prone to being degraded over time, resulting in irreparable loss. Optical Music Recognition (OMR) is the field within computer vision that has aimed, for decades, to try to mitigate this issue by finding means of which the music contained within music scores can be preserved in a machine-readable, replicable format. Even so, OMR has presented itself as a field of study with a considerable barrier of entry, not only because of the need for knowledge within the world of music theory, but also a myriad of complex algorithms concerned with the multiple steps of the traditional OMR pipeline. The recent embrace of deep learning algorithms, able generalize learning of image features regardless of noise and distortions has made great strides to simplify the OMR pipeline. Due to the positioning of music objects as well as their features needing to be accounted for to properly classify them, this study has tackled OMR through an object detection approach. It makes use of Ultralytics’ YOLOv8 algorithm and has managed to detect, classify and reconstruct the digitally written, monophonic music scores contained within the "Printed Images of Music Staves" (PrIMuS) dataset with mAP50-95 scores above 85%.