Binarization of music score images using line width transform

Although the original Gaussian Mixture Markov Random Field model can generate good binarization results for scene text images, it still has some issues needed to be solved in case of music score images. The difficulty is the ineffective seeding algorithm when it is applied to music score images which consist of thin lines, and isolated and complex background regions. A wrong seeding makes the false positive and false negative in foreground and background labelling. In this paper, a new adaptive model for the binarization of complex background music score image is proposed. We suggest a line width transform based seeding method for a better GMMs initialization of foreground and background color distribution in music score image. The result is the better binarization with cleaner background and clearer foreground compared to previous binarization techniques.