[kineticfocus]

"...divides a single image into many different images, then compresses them separately." Wouldn't it be better to go the other way and backpack-algo the pages into a consolidated texture atlas based on edge similarity, and then run compression on the result?

[dunham]

From what I understand, it segments the image into text and a background image, does something like JBIG2 on the text (building a global dictionary of glyphs and placing them on the page), and does a lossy compression on the background image.

There aren't a lot of viewers out there for DjVu and the encoding side is patent encumbered, so I'm not interested in the format.

You can get pretty close with JBIG2+Jpeg2k in a PDF file, I believe archive.org does this, but I don't know of an open source encoder that does it and sometimes PDF viewers don't decode jbig2/jpg2k efficiently.