Hola

Solo disponible en BuenasTareas
  • Páginas : 12 (2815 palabras )
  • Descarga(s) : 0
  • Publicado : 5 de diciembre de 2010
Leer documento completo
Vista previa del texto
A COMPARATIVE SURVEY OF IMAGE BINARISATION ALGORITHMS FOR OPTICAL RECOGNITION ON DEGRADED MUSICAL SOURCES
John Ashley Burgoyne Laurent Pugin Greg Eustace Ichiro Fujinaga Centre for Interdisciplinary Research in Music and Media Technology Schulich School of Music of McGill University Montr´ al, Qu´ bec, Canada H3A 1E3 e e {ashley,laurent,greg,ich}@music.mcgill.ca
ABSTRACT Binarisation ofgreyscale images is a critical step in optical music recognition (OMR) preprocessing. Binarising music documents is particularly challenging because of the nature of music notation, even more so when the sources are degraded, e.g., with ink bleed-through from the other side of the page. This paper presents a comparative evaluation of 25 binarisation algorithms tested on a set of 100 music pages. Areal-world OMR infrastructure for early music (Aruspix) was used to perform an objective, goaldirected evaluation of the algorithms’ performance. Our results differ significantly from the ones obtained in studies on non-music documents, which highlights the importance of developing tools specific to our community. 1 INTRODUCTION Binarising music documents, that is separating the foreground from thebackground in order to prepare for other tasks such as optical music recognition (OMR), is much more challenging than binarising text documents. In text documents, the letters are all of approximately the same size and are regularly and uniformly distributed throughout the page. Music symbols, on the other hand, exhibit a wide range of sizes and markedly uneven distribution: they are clusteredaround musical staves. Large black areas, such as note heads, are conducive to ink accumulation during printing, which often results in strong bleedthrough (elements from the verso visible through the paper), especially for early sources. Large blank areas without foreground elements can disturb some binarisation techniques because bleed-through is often considered to be foreground. We will show inthis paper that because of these conditions, the most widely used binarisation methods fail to produce suitable results for OMR. Evaluating the performance of binarisation algorithms is a difficult task. Very often, due to a lack of any evaluation infrastructure, researchers use subjective approaches, e.g., marking output as “better”, “same” or “worst” [3, 4]. When the binarisation is performed forthe purpose of further image processing tasks, such as optical character c 2007 Austrian Computer Society (OCG). recognition (OCR) or OMR, it makes more sense to use an objective evaluation. Evaluating the algorithms within the context of a real-world application enables goal-directed evaluation, which rates a binarisation algorithm on its ability to improve the post-binarisation task [10].Furthermore, it has been shown that when document images have graphical particularities like music documents do, the use of goal-directed evaluation can lead to significant performance improvements [7]. 2 METHODS For our experiments, we used Aruspix, a software application for OMR on early music prints [7]. We selected five 16th-century music books (RISM 1520-2, 1532-10, 15385, M-0579 and M-0582 [8])that suffer from severe degradation and transcribed 20 pages from each (100 total) to obtain ground-truth data for the evaluation. We tested 25 different binarisation algorithms over a range of parameters, which resulted in a set of 8,000 images. The images were deskewed and normalised to a consistent staff height (100 pixels) by Aruspix before applying the binarisation algorithm, and afterbinarisation, Aruspix was used again for the OMR evaluation. Binarisation methods can be categorised according to differences in the criteria used for thresholding. Sezgin and Sankur have proposed a taxonomy of thresholding techniques, including those based on the shape of the greyvalue histogram, measurement-space clustering, image entropy, connected-component attributes, spatial correlation, and the...
tracking img