Scanned scores are black ink on white paper stored as 8-bit greyscale or
RGB, which costs several times what the same page costs as a bilevel
image. Across an 11-song corpus this is 27.6 MB to 7.7 MB; Engel's
bundle goes from 5997 KB to 2092 KB with byte-identical slices, since
only the archived copy changes.
Three approaches were measured and discarded first, which is worth
recording because two of them are the obvious ones. Converting RGB to
greyscale and re-encoding makes these files 20-86% LARGER: the source
JPEGs are already near 0.7 bits per pixel, so re-encoding adds
generation loss and spends more bits than the original did, and dropping
chroma recovers nothing because JPEG already subsamples it. Lossless
structural optimisation gains 0.1%, because images are 99% of every file
and there are no duplicates. Downsampling works but 300 DPI is print
resolution, and the PDF exists to be printed.
Two failure modes were found by looking at output rather than at byte
counts, and both are now refused:
- A scan at ~115 DPI came back with broken staff lines. Guarded on
resolution as the image is *placed on the page*, so a tiled scan with
126 small images still qualifies where a pixel count would reject it.
- Cover artwork was flattened to grey. Guarded on chroma: artwork
measures 44% off-grey against 3% for sensor tint on a greyscale scan.
The first threshold of 2% was a false positive that cost 685 KB on one
song for nothing; 10% sits in the gap with room either side.
Exposed as a button rather than a checkbox. It reports what it skipped
and why, and shows a before/after crop, because the failure it can
produce is obvious at a glance and invisible in a size figure. Off by
default: this is lossy on the copy kept for printing.
