This is the supplementary material with automatic stave discovery, binarization and staff removal results for
the algorithms presented in the paper. Always we use 5-groups of staff lines for staves and the parameters from the paper.
These are examples meant to address steps and methods described in the paper.
pdf
- the original input image.
Note that all the images included were reduced to a lower resolution (the images were captured at >100Megapixels resolution)
for reasons of space and to prove that our methods work well in these conditions. The images were cropped in order to do not reveal the origin (keep anonymous).
Moreover, no other color/image processing methods were employed to normalize, enhance, adapt the input images.
Staves
- the discovered staves and the segmentation are overlaid over the original image.
With green segments are marked the musical staves as segmented. With blue segments are marked the parts that belong to background (not staves).
Note that we employ a single pass (iteration) for stave discovery over all the images from this supplementary material.
Moreover, at this point the results do not contain accurate staff detection.
Stave content
- the discovered stave segments as corresponding to the original image, with the background suppressed to white.
Binarization
- binary image, result of binarization process. Note that due to the fact that we use a global threshold, there are cases where parts of the scores are lost.
Staff removal
- the clean musical symbols after applying stave discovery, stave segmentation, accurate staff detection, and staff removal steps. No other post-processing is employed
Score 1
a quite clean input image
Original
Staves
Stave content
Binarization
Staff removal
Score 2
an interesting case - no musical symbols, just staves
Original
Staves
Stave content
Binarization
Staff removal
Score 3
two voices, a middle stave with text, below the 6th stave there is a staffline segment that survives the staff removal process because it is outside the stave -- we use the methods like described in the paper
Original
Staves
Stave content
Binarization
Staff removal
Score 4
big illustration, interruption with lyrics, strong shadows, small stave segment. While picked by the detection step, the stave going through lyrics is labeled correctly as non-stave / background
Original
Staves
Stave content
Binarization
Staff removal
Score 5
Original
Staves
Stave content
Binarization
Staff removal
Score 6
a dense score, the last 2 staves are with different spacing than the first 9 staves. The method picks the last stave, but labels it as background since do not fit the expected model built for the first 9 staves.
To get staves with different staffline thicknesses and spacings a solution is to repeat the stave discovery process after masking the previously detected and labeled staves.
Original
Staves
Stave content
Binarization
Staff removal
Score 7
nice illustrations, two voices, shaded areas where the global threshold fails to clearly separate musical content from background
Original
Staves
Stave content
Binarization
Staff removal
Score 8
Original
Staves
Stave content
Binarization
Staff removal
Score 9
due to the fact that (here) we cut according to the stave segment, we loose musical symbols that are next to the staves
Original
Staves
Stave content
Binarization
Staff removal
Score 10
due to the fact that we cut according to the stave segment, we might loose musical symbols that are next to the staves
Original
Staves
Stave content
Binarization
Staff removal
Score 11
a dense score, the last stave is not correctly segmented (with multiple iterations for stave discovery, this would not happen), due to strong shadows and use of global threshold the end of some staves are missed
Original
Staves
Stave content
Binarization
Staff removal
Score 12
we run 5-groups of stafflines over 4-groups of stafflines scores and surprisingly the results are quite good, despite wrongfully labeling some stave segments (upper right corner)
Original
Staves
Stave content
Binarization
Staff removal
Score 13
with a 5-groups of stafflines system over 4-groups of stafflines scores and very good results
Original
Staves
Stave content
Binarization
Staff removal
Score 14
an interesting case - back-shadows automatically picked by the system. The first stave was too saturated wrt the others and got rejected by the stave discovery process.
However, in applications the user can limit the values range from where the automatic procedure picks the threshold, thus this case to do not happen.
Original
Staves
Stave content
Binarization
Staff removal
Score 15
three voices
Original
Staves
Stave content
Binarization
Staff removal
Score 16
Original
Staves
Stave content
Binarization
Staff removal
Score 17
back shadows brought to surface (middle left and top left corner), staff removal fails in removing the short middle stave from the right side
Original
Staves
Stave content
Binarization
Staff removal
Score 18
three voices
Original
Staves
Stave content
Binarization
Staff removal
Score 19
three voices, two short staves are not detected after the one pass/iteration of the automatic stave discovery procedure. After the second iteration, these are properly detected.
Original
Staves
Stave content
Binarization
Staff removal
Score 20
small background segment in middle of a stave
Original
Staves
Stave content
Binarization
Staff removal
Score 21
symbols out or at the end of the staves that are missed. the cropping follows the stave segments
Original
Staves
Stave content
Binarization
Staff removal
Score 22
large illustrations, good results
Original
Staves
Stave content
Binarization
Staff removal
Score 23
dense score with symbols out or at the end of the staves that are missed. the cropping follows the stave segments
Original
Staves
Stave content
Binarization
Staff removal
Score 24
dense score with symbols out or at the end of the staves that are missed. the cropping follows the stave segments
Original
Staves
Stave content
Binarization
Staff removal
Score 25
good results, however the symbols are strongly darkened