Marshall Bockrath

One last self-reply with a few additional findings: (1) I realized I'd not used the threshold I'd intended when binarizing the page images for Programming the Commodore 64. Using a much lower threshold and generating lighter glyphs eliminated most of the regular/bold confusion across the board, but did not fix all of the [Il1] confusion. (2) I found a bit more context and detail on jbig2enc and the classification algorithms, in the paper "Google Books: Making the public domain universally accessible"...

sets pages-per-dict to ∞ That was my intent, but I didn't consider that this might have bad effects on the final output. jbig2 only operates on and builds a dictionary for the entire set of pages, so I was trying to make sure that aspect of the comparison was apples-to-apples. I just ran a set of all the minidjvu/-mod runs with -p 10, and the -p 0 results were an average of 15% smaller, which should only improve further with total page count. Aren't you miss a 8 in timing of a first 3 rows here?...

Are you using GUI tool to launch them? 😂 I've got the raw data, scripts, and my results here for a relatively rigorous comparison: http://files.platypope.org/compare.tar.gz The input data is the first 103 pages of the book Programming the Commodore 64 , derived from the Internet Archive's scans. (I didn't say it was a good technical book heh). It isn't as typographically complex as a math text, but it does include multiple typefaces with multiple variations. Critically, in the main body text typeface...

Are you using GUI tool to launch them? 😂 I've got the raw data, scripts, and my results here for a relatively rigorous comparison: http://files.platypope.org/compare.tar.gz The input data is the first 103 pages of the book Programming the Commodore 64 , derived from the Internet Archive's scans. (I didn't say it was a good technical book heh). It isn't as typographically complex as a math text, but it does include multiple typefaces with multiple variations. Critically, in the main body text typeface...

Sorry, to self-reply, but I spoke too soon. The too-good-to-be-true "lossless" output from minidjvu was too good and isn't lossless -- not at all. It didn't make the specific incorrect substitutions I'd been spot-checking, but when I continued reading, I started noticing quite a few, mostly of glyphs in the regular typeface being replaced with glyphs in the bold typeface. It's too late for me to try to understand what happened, but I'm confident my command-line specifies neither -lossy nor -match,...

Both. For the two books I'm working with right now, each technical books with ~650 pages each, I didn't see a big difference between minidjvu and minidjvu-mod in either speed or apparent accuracy. (Yes, minidjvu-mod can run multiple dictionary page sets in parallel, but I'm talking about single-threaded performance of the similarity algorithm.) (Also, I realized I completely neglected to link to the actual Leptonica classifier, which may be found in its jbclass.c source file.) Performance -- the...

Both. For the two books I'm working with right now, each technical books with ~650 pages each, I didn't see a big difference between minidjvu and minidjvu-mod in either speed or apparent accuracy. (Yes, minidjvu-mod can run multiple dictionary page sets in parallel, but I'm talking about single-threaded performance of the similarity algorithm.) (Also, I realized I completely neglected to to the actual Leptonica classifier, which may be found in its jbclass.c source file.) Performance -- the main...

Both. For the two books I'm working with right now, each technical books with ~650 pages each, I didn't see a big difference between minidjvu and minidjvu-mod in either speed or apparent accuracy. (Yes, minidjvu-mod can run multiple dictionary page sets in parallel, but I'm talking about single-threaded performance of the similarity algorithm.) (Also, I realized I completely neglected to to the actual Leptonica classifier, which may be found in its jbclass.c source file.) Performance -- the main...