Sort every pixel on a scanned page by how dark it is, and they pile up
at the two ends. There is ink, and there is paper, and almost nothing
in between:
The 254 shades in the middle are paper texture, sensor noise and the
soft halo JPEG leaves around every note stem. Throwing them away costs
nothing, because there was nothing there — and it means a pixel no
longer needs to record which of 256 shades it is, only whether
it is ink:
8 bits per pixel — one of 256 shades
1 bit per pixel — ink, or not ink
That alone is eight times smaller. Group 4 then squeezes the long
runs of untouched paper — most of a page of music is empty —
and that is where the rest of the 15–20× comes from.