Enhanced Bleedthrough Correction for Early Music Documents with Recto-Verso Registration

Ink bleedthrough is common problem in early music documents. Even when such bleedthrough does not pose problems for human perception, it can inhibit the performance of optical music recognition (OMR). One way to reduce the amount of bleedthrough is to take into account what is printed on the reverse of the page. In order to do so, the reverse of the page must be registered to match the front of the page on a pixel-by-pixel basis. This paper describes our approach to registering scanned early music scores as well as our modifications to two robust binarization approaches to take into account bleedthrough and the information available from the registration process. We determined that although the information from registration itself often makes little difference in recognition performance, other modifications to binarization algorithms for correcting bleedthrough can yield dramatic increases in OMR results. 1 MOTIVATION AND BACKGROUND 1.1 Fostering Interdisciplinary Research with OMR “We stand at a moment of opportunity,” opened Nicholas Cook at his invited talk for ISMIR 2005 in London. The opportunity is for historical musicologists and music information scientists to work together and revitalize the subdiscipline of computer-assisted empirical musicology [6]. This subdiscipline began in the 1960s [15], and although it has developed into a thriving discipline in several regions, it is largely moribund in North America, where a significant amount of other musicological research takes place. While Cook assigned musicologists a great deal of the responsibility for realizing this moment of interdisciplinary opportunity, he challenged researchers in music information retrieval to create large databases of the “highly reduced data”—e.g., scores—upon which musicological research relies. Such electronic databases are especially important for those older documents that are available in only a limited number of locations and for which archivists often restrict physical access, making it difficult to engage in large-scale comparative research. Entering musical sources into such databases by hand is highly labor-intensive, which renders music digitization projects prohibitively costly for most institutions [3]. Optical music recognition (OMR), the musical analog to optical character recognition (OCR), is the most practical means by which to create such databases. The potential for optical recognition to transform research approaches has already been demonstrated by the recent explosion in the number of searchable electronic texts available for older books and journal articles, (e.g., JSTOR 1 ). When treating historical documents, however, OMR and OCR systems struggle with various types of document degradation, including ink bleedthrough from the reverse side of the page [1]. Because these systems rely on the ability to distinguish foreground (ink) from background (paper), the darker the bleedthrough, the more likely it is that the bleedthrough will be classified as foreground and thereby corrupt the recognition process. 1.2 Binarization and Bleedthrough Binarization is the common name for separating an image into its foreground and background. It is notoriously difficult to evaluate, leading most authors to resort to coarse subjective distinctions such as “better,” “same,” or “worse,” e.g., [12]. As binarization is typically a preprocessing step in an automated image-processing pipeline with some other goal, e.g., OMR, one can use the evaluation metric for the ultimate goal, e.g., minimizing the amount of time necessary for a human editor to correct recognition errors, to evaluate the quality of the binarization algorithm. This paper uses a similar method to that described in our recently-published survey of binarization algorithms for music documents [4] to evaluate extensions to the algorithms we found to be bestperforming. Although there has been some previous work on binarization in the presence of bleedthrough that focuses specifically on historical documents [9, 10], a larger body of work 1 http://www.jstor.org/