Hand back what the codec read, and the mask beside it

[tannevaled]

Images returned what the page draws: the picture with its mask applied. That made it useless for the job it was added for — and worse than useless, because it was wrong in a way that looked like a finding.

Judged against pdfimages over a corpus of scanned pages, 21 of 22 JPEG 2000 pictures came out differing, at a median of 85% of their pixels.

That reading was impossible on its face: the same pages, compared whole against the same renderer, differ by a median of 0.0000. A picture cannot be 85% wrong on a page that is perfect.

Eleven of the twenty-two carry an /SMask. We applied it and pdfimages did not, so the two sides were not being asked the same question.

What changes

Images now hands back what each codec produced, with the mask beside it as its own entry named for the key that named it (I/SMask) — which is what pdfimages does, and the only way the two can be compared. A mask is a picture in its own right, stored in its own filter, and very often the one that is wrong: JBIG2 is almost never a page's content and almost always its mask.

A picture whose mask cannot be read now comes back too. Page still declines to draw that one, because how much of it shows is unknown; the codec read it, and this is about the codec. Both halves are asserted in one test, because it would be easy to move the first and lose the second.

decodeImage keeps its meaning — what the page draws — and is now that composition written down: decodeBase, then the mask.

100% statement coverage, go vet and -race clean, nine cross-compile targets.