Complex layout analysis
Advancing well on complex layout analysis. Here is what we showed for the trainable baseline detection for complex Arabic documents at the last HIP@ICDAR in September 2019 (see publications). It is very successful on documents of varying line orientation.
With a succeeding multidimensional seamcarve we then isolate the polygon around a baseline as is demonstrated in the second picture, a Hebrew manuscript from Geneva in the Tikkoun Sofrim project. Implementation in kraken / eScriptorium should be out any day now.