Fixed:
* `recognize`: regression from changed network initialization
* `recognize`: also load uncompressed models (Python 3 port)
* `re/segment`: avoid creating invalid polygon coordinates
* ocrolib scale estimation: make DPI-dependent, add fallback for empty/noise pages
* ocrolib morphology: avoid rounding artifacts and asymmetry
Changed:
* ocrolib / all ocropy processors: boost performance via OpenCV/PIL intead of SciPy
* `binarize`: expose `threshold` parameter
* `re/segment`: require images to be binarized already instead of ad-hoc binarization
* `re/segment`: much faster line segmentation, better separation of neighbouring lines
* `segment`: much more robust fg h/v-line and bg column detection, new image detection
* `segment`: add `AlternativeImage` with h/v-line or image non-text clipped to background
* `segment`: rewrite of region aggregation via hybrid recursive X-Y cut
* `segment`: also annotate detected lines (at detected regions) after page segmentation
* `segment`: expose many new parameters
* `segment`: add all new lines/regions in proper (but only top-down left-right) reading order
* `segment`: add `table` level (like page segmentation, but horizontal-first split strategy)
* `segment`: also recurse into table cells for `region` level
* `segment`: incremental annotation (ignore and re-order existing text/image/separator regions)
* `profile`: remove this CLI, but keep integrated via API in new pure OCR-D CLI `postcorrect`
* `wer`, `training`: remove these CLI
* `postcorrect`, `align`: update to latest JAR
* remove non-OCR-D scripts from installation
* add `uninstall` target
* update documentation (esp. training, testing and postcorrection)
* improve/extend automatic tests