Identification and Cross-Document Alignment of Measures in Music Score Images

In the course of editing musical works, musicologists regularly compare multiple sources of the same musical piece, such as composers' autographs, handwritten copies, and various prints. For efficient comparison, cross-source navigation is essential, enabling to quickly jump back and forth between multiple sources without losing the current musical position. In practice, measures are first annotated by hand in the individual source images and then related to each other. Our approach automates this time-consuming and error-prone process with the help of deep learning. For this purpose, we train a neural network that automatically finds bounding boxes of all measures in images. A second network is trained to compute the similarity between two measures to determine if they have the same musical content and should, therefore, be linked for navigation. Sequences of outputs from the second network are matched using Dynamic Time Warping to provide the final proposal of measure relationships, so-called concordances. In addition to cross-source navigation, the results can be used to spot structural differences across the sources which are essential for editorial work, so that musicologists can focus more on analytical tasks