Image Quality Estimation for Multi-Score OMR

Optical music recognition (OMR) is the recognition of images of musical scores. Recent research has suggested aligning the results of OMR from multiple scores of the same work (multi-score OMR, MS-OMR) to improve accuracy. As a simpler alternative, we have developed features which predict the quality of a given score, allowing us to select the highest-quality score to use for OMR. Furthermore, quality may be used to weight each score in an alignment, which should improve existing systems’ robustness. Using commercial OMR software on a test set of MIDI recordings and multiple corresponding scores, our predicted OMR accuracy is weakly but significantly correlated with the true accuracy. Improved features should be able to produce highly consistent results.