Sequential Next-Symbol Prediction for Optical Music Recognition

Optical Music Recognition is the research field that investigates how to computationally read music notation from document images. State-of-the-art technologies, based on Convolutional Recurrent Neural Networks, typically follow an end-to-end approach that operates at the staff level; i.e., a single stage for completely processing the image of a single staff and retrieving the series of symbols that appear therein. This type of models demands a training set of sufficient size; however, the existence of many music manuscripts of reduced size questions the usefulness of this framework. In order to address such a drawback, we propose a sequential classification-based approach for music documents that processes sequentially the staff image. This is achieved by predicting, in the proper reading order, the symbol locations and their corresponding music-notation labels. Our experimental results report a noticeable improvement over previous attempts in scenarios of limited ground truth (for instance, decreasing the Symbol Error Rate from 70% to 37% with just 80 training staves), while still attaining a competitive performance as the training set size increases.