Collation - Overview

Development

It was necessary to develop a specific solution for the optical recognition of early printed sources, since the comparison of different editions of a work cannot always be performed by a straightforward superimposition of images. One of the principal reasons is that the layouts can be quite different from one edition to another.

However, the optical recognition of early music prints is not an easy task. Existing recognition software is usually unable to cope with the amount of noise present in an image scanned from microfilm, or with the curvature of staff lines and the fact that the pages are rarely straight. The musical characters are unusual and often touch other characters, making them difficult to distinguish.

Processes used

Aruspix operates in several phases, by first cleaning the original image and correcting page skew, and then by pre-classifying certain elements. The result of these pre-treatment and pre-classification phases is an image that can then be digitally recognized for musical content. To do this, Aruspix must first detect the curvature of the staff lines, for which we developed a specific solution.

Aruspix then uses two independent musical models. The typographic model uses Hidden Markov Models (HMM) and must be trained from examples. The stochastic model validates the results, based on different fields of information, but without imposing constraining rules that would limit the efficiency of the typographic model.

Finally, Aruspix displays the recognition results in an integrated music editor that allows them to be superimposed on the original image and any errors to be corrected quickly and easily.

Results achieved

The specific solutions developed within Aruspix have enabled fairly high recognition rates to be achieved (around 97-98%), even when the original images contained a great deal of noise or were very distorted. Now that the typographic model has been sufficiently trained, the entire process of pre-treatment, pre-classification and recognition takes just a few seconds for a typical page, and batch processing and recognition has been performed on entire books in less than half an hour.

Future prospects

The comparison of the resulting recognized images remains to be implemented, as well as recognition of the text elements that were isolated in the pre-classification phase. Once this is done, Aruspix will constitute a complete, self-contained recognition and comparison software application that can then be extended to other types of documents. The more the software is used, the better the typographic and stochastic models will become, meaning that a database of music models can be built up, in addition to the prospect of a database of the musical content of scanned and recognized sources.