✨ Use sbb_textline_detector to segment lines

Dockerfile

@@ -17,6 +17,8 @@ RUN apt-get update && \
       cmake libgif-dev libjpeg-dev libpng-dev libtiff-dev zlib1g-dev \
 # For clstm on Ubuntu 19.04:
       swig libeigen3-dev libpng-dev libprotobuf-dev \
+# For cv2:
+      libsm6 libxrender1 \
 # XML utils
       libxml2-utils \
       xmlstarlet \
@@ -53,6 +55,11 @@ COPY data/tesseract-models/GT4HistOCR/GT4HistOCR_2000000.traineddata $TESSDATA_P
 RUN tesseract --list-langs
 
 
+# Copy over sbb_textline_detector
+COPY vendor vendor
+COPY data/textline_detection /var/lib/textline_detection
+
+
 COPY requirements.txt /tmp
 RUN pip3 install --no-cache-dir -r /tmp/requirements.txt
 

build

@@ -7,6 +7,7 @@ set -e
   git annex upgrade
   git annex get calamari-models/GT4HistOCR/*.ckpt*
   git annex get tesseract-models/GT4HistOCR/*.traineddata
+  git annex get textline_detection/*.h5
 )
 
 docker build -t my_ocrd_workflow .

data

@@ -1 +1 @@
-Subproject commit eb7412a1efbcba53567ec37237732e96e839dbe8
+Subproject commit bcc1aec082cb81c29668ffef3d04c51eaa866b5c

my_ocrd_workflow

@@ -59,7 +59,7 @@ do_fontident() {
   #     any DEFAULT, yet -I DEFAULT seems to work for ocrd-typegroups-classifier
 }
 
-do_linesegmentation() {
+do_linesegmentation_tesserocr() {
   # Segment the lines in the binarized images
 
   remove_filegrp OCR-D-SEG-REGION mets.xml
@@ -76,6 +76,16 @@ do_linesegmentation() {
   # XXX compare ocrd-tesserocr-segment* vs tesseract native
 }
 
+do_linesegmentation_sbb() {
+  # Segment the lines in the images
+
+  remove_filegrp OCR-D-SEG-REGION mets.xml
+  remove_filegrp OCR-D-SEG-LINE mets.xml
+  ocrd_sbb_textline_detector -l $LOG_LEVEL \
+    -m mets.xml -I OCR-D-IMG -O OCR-D-SEG-LINE \
+    -p '{"model": "/var/lib/textline_detection"}'
+}
+
 do_ocr() {
   # Perform OCR on the segmented lines
 
@@ -123,16 +133,22 @@ page_fix_image_references() {
   done
 }
 
-page_workaround_remove_conf() {
+page_fix_image_references_to_bin() {
-  # XXX Work around https://github.com/OCR-D/core/issues/269
+  # Make image references point to the binarized images
+  # XXX This is a hack, it is probably better to use alternative images in ocrd_calamari
 
   filegrp=$1
 
   local file
   for file in `ocrd workspace find -G $filegrp`; do
-    xmlstarlet ed --inplace \
+    # Arrays with filenames to the images
-    -N 'page=http://schema.primaresearch.org/PAGE/gts/pagecontent/2019-07-15' \
+    imgs=(`ocrd workspace find -G OCR-D-IMG`)
-    -d '//page:TextEquiv/@conf' $file
+    imgs_bin=(`ocrd workspace find -G OCR-D-IMG-BIN`)
+
+    # Change all image references to point to the corresponding binarized image
+    for i in ${!imgs[@]}; do
+      sed -i "s!imageFilename=.${imgs[$i]}.!imageFilename=\"${imgs_bin[$i]}\"!g" $file
+    done
   done
 }
 
@@ -146,6 +162,14 @@ page_downgrade_to_2018() {
   done
 }
 
+page_upgrade_to_2019() {
+  filegrp=$1
+
+  local file
+  for file in `ocrd workspace find -G $filegrp`; do
+    sed -i 's#pagecontent/[0-9-]*#pagecontent/2019-07-15#g' $file
+  done
+}
 
 pip3 list
 
@@ -158,9 +182,11 @@ do_binarization
 do_validate
 
 
-do_linesegmentation
+do_linesegmentation_sbb
-page_validate_xml           OCR-D-SEG-REGION
+page_fix_image_references_to_bin OCR-D-SEG-LINE
-page_validate_xml           OCR-D-SEG-LINE
+page_upgrade_to_2019             OCR-D-SEG-LINE
+page_validate_xml                OCR-D-SEG-REGION
+page_validate_xml                OCR-D-SEG-LINE
 do_validate
 
 

requirements.txt

@@ -1,4 +1,6 @@
-Pillow==5.4.1  # See https://github.com/OCR-D/core/issues/325
+tensorflow-gpu < 2.0  # Needed for sbb_text_linedetector
+Pillow==5.4.1         # See https://github.com/OCR-D/core/issues/325
+
 ocrd >= 1.0.0b19
 
 https://github.com/mikegerber/ocrd_typegroups_classifier/archive/fix/pass-down-page-id.tar.gz  # XXX git+https://github.com/seuretm/ocrd_typegroups_classifier.git
@@ -9,4 +11,6 @@ ocrd_tesserocr
 
 https://github.com/mikegerber/ocrd_calamari/archive/3e8c1ac.tar.gz
 
+vendor/sbb_textline_detector-b1663f7.tar
+
 https://github.com/qurator-spk/dinglehopper/archive/c305539.tar.gz