Binarization of music score with complex background by deep convolutional neural networks

Abstract

Binarization is an important step for most of document analysis systems. Regarding music score images with a complex background, the existence of background clutters with a variety of shapes and colors creates many challenges for the binarization. This paper presents a model for binarization of the complex background music score images by fusion of deep convolutional neural networks. Our model is directly trained from image regions using pixel values as inputs and the binary ground truth as labels. By utilizing the generalization capability of the residual network backbone and useful feature learning ability of dense layer, the proposed network structures can differentiate foreground pixels from background clutters, minimize the possibility of overfitting phenomenon and thus can deal with complex background noises appearing in the music score images. Comparing to traditional algorithms, binary images generated by our method have a cleaner background and better-preserved strokes. The experiments with captured and synthetic music score images show promising results compared to existing methods.

Music Score Recognition and Composition Application Based on Deep Learning

“…It is possible for the statistical transformation approach to fail when the spectral line has been deformed and does not conform to a rigid straight line shape; however, this is not likely to happen in most cases. Local details are easily influenced by noise interference, which might be difficult to detect [30]. When the level of interference reaches a specific threshold, the system will be forced to contend with the issue of insufficient local information collecting and no general direction.…”