Using jbig2enc in conjunction with minidjvu/minidjvu-mod.
The jbig2enc implementation uses the leptonica classifier for lossy compression, which is unsatisfactory of:
Kjbig2 = Kpack / (1.0 - Kerror)
Kpack = size_jbig2 / size_pbm
Kerror = sqrt(1.0 - part_no_error * part_no_error)
part_no_error = (full - error) / full
To overcome poor compression performance and classification errors, a recipe will be shown for using the third-party minidjvu-mod classifier with increasing the threshold of the leptonica classifier to eliminate classification errors.