✨ Allow controlling of output hierarchy level, e.g. only line, not words+glyphs

README.md

@@ -13,10 +13,11 @@ This offers a OCR-D compliant workspace processor for some of the functionality
 This processor only operates on the text line level and so needs a line segmentation (and by extension a binarized 
 image) as its input.
 
-In addition to the line text it also outputs glyph segmentation including
+In addition to the line text it may also output word and glyph segmentation
-per-glyph confidence values and per-glyph alternative predictions as provided by
+including per-glyph confidence values and per-glyph alternative predictions as
-the Calamari OCR engine. Note that while Calamari does not provide word
+provided by the Calamari OCR engine, using a `textequiv_level` of `word` or
-segmentation, this processor produces word segmentation inferred from text
+`glyph`. Note that while Calamari does not provide word segmentation, this
+processor produces word segmentation inferred from text
 segmentation and the glyph positions. The provided glyph and word segmentation
 can be used for text extraction and highlighting, but is probably not useful for
 further image-based processing.
@@ -53,7 +54,8 @@ ocrd-calamari-recognize -p test-parameters.json -m mets.xml -I OCR-D-SEG-LINE -O
 With `test-parameters.json`:
 ~~~
 {
-    "checkpoint": "/path/to/some/trained/models/*.ckpt.json"
+    "checkpoint": "/path/to/some/trained/models/*.ckpt.json",
+    "textequiv_level": "line",
 }
 ~~~
 

ocrd_calamari/ocrd-tool.json

@@ -25,6 +25,12 @@
         "voter": {
           "description": "The voting algorithm to use",
           "type": "string", "default": "confidence_voter_default_ctc"
+        },
+        "textequiv_level": {
+          "type": "string",
+          "enum": ["line", "word", "glyph"],
+          "default": "line",
+          "description": "Deepest PAGE XML hierarchy level to include TextEquiv results for"
         }
       }
     }

ocrd_calamari/recognize.py

@@ -120,46 +120,48 @@ class CalamariRecognize(Processor):
                                 spaces = (c == ' ')
                         yield word
 
-                    word_no = 0
+                    if self.parameter['textequiv_level'] in ['word', 'glyph']:
-                    i = 0
+                        word_no = 0
+                        i = 0
 
-                    for word_text in _words(prediction.sentence):
+                        for word_text in _words(prediction.sentence):
-                        word_length = len(word_text)
+                            word_length = len(word_text)
-                        if not all(c == ' ' for c in word_text):
+                            if not all(c == ' ' for c in word_text):
-                            word_positions = prediction.positions[i:i+word_length]
+                                word_positions = prediction.positions[i:i+word_length]
-                            word_start = word_positions[0].global_start
+                                word_start = word_positions[0].global_start
-                            word_end = word_positions[-1].global_end
+                                word_end = word_positions[-1].global_end
 
-                            polygon = polygon_from_x0y0x1y1([word_start, 0, word_end, line_image.height])
+                                polygon = polygon_from_x0y0x1y1([word_start, 0, word_end, line_image.height])
-                            points = points_from_polygon(coordinates_for_segment(polygon, None, line_coords))
-                            # XXX Crop to line polygon?
-
-                            word = WordType(id='%s_word%04d' % (line.id, word_no), Coords=CoordsType(points))
-                            word.add_TextEquiv(TextEquivType(Unicode=word_text))
-
-                            for glyph_no, p in enumerate(word_positions):
-                                glyph_start = p.global_start
-                                glyph_end = p.global_end
-
-                                polygon = polygon_from_x0y0x1y1([glyph_start, 0, glyph_end, line_image.height])
                                 points = points_from_polygon(coordinates_for_segment(polygon, None, line_coords))
+                                # XXX Crop to line polygon?
 
-                                glyph = GlyphType(id='%s_glyph%04d' % (word.id, glyph_no), Coords=CoordsType(points))
+                                word = WordType(id='%s_word%04d' % (line.id, word_no), Coords=CoordsType(points))
+                                word.add_TextEquiv(TextEquivType(Unicode=word_text))
 
-                                chars = sorted(p.chars, key=lambda k: k.probability, reverse=True)
+                                if self.parameter['textequiv_level'] == 'glyph':
-                                char_index = 1  # Must start with 1, see https://ocr-d.github.io/page#multiple-textequivs
+                                    for glyph_no, p in enumerate(word_positions):
-                                for char in chars:
+                                        glyph_start = p.global_start
-                                    if char.char:
+                                        glyph_end = p.global_end
-                                        glyph.add_TextEquiv(TextEquivType(Unicode=char.char, index=char_index, conf=char.probability))
-                                        char_index += 1
-                                        # XXX Note that omission probabilities are not normalized?!
 
-                                word.add_Glyph(glyph)
+                                        polygon = polygon_from_x0y0x1y1([glyph_start, 0, glyph_end, line_image.height])
+                                        points = points_from_polygon(coordinates_for_segment(polygon, None, line_coords))
 
-                            line.add_Word(word)
+                                        glyph = GlyphType(id='%s_glyph%04d' % (word.id, glyph_no), Coords=CoordsType(points))
-                            word_no += 1
 
-                        i += word_length
+                                        chars = sorted(p.chars, key=lambda k: k.probability, reverse=True)
+                                        char_index = 1  # Must start with 1, see https://ocr-d.github.io/page#multiple-textequivs
+                                        for char in chars:
+                                            if char.char:
+                                                glyph.add_TextEquiv(TextEquivType(Unicode=char.char, index=char_index, conf=char.probability))
+                                                char_index += 1
+                                                # XXX Note that omission probabilities are not normalized?!
+
+                                        word.add_Glyph(glyph)
+
+                                line.add_Word(word)
+                                word_no += 1
+
+                            i += word_length
 
 
             _page_update_higher_textequiv_levels('line', pcgts)

test/test_recognize.py

@@ -87,6 +87,7 @@ def test_word_segmentation(workspace):
         output_file_grp="OCR-D-OCR-CALAMARI",
         parameter={
             "checkpoint": CHECKPOINT,
+            "textequiv_level": "word",   # Note that we're going down to word level here
         }
     ).process()
     workspace.save_mets()
@@ -106,5 +107,30 @@ def test_word_segmentation(workspace):
     line_text = line.xpath("pc:TextEquiv/pc:Unicode", namespaces=NSMAP)[0].text
     assert words_text == line_text
 
+    # For extra measure, check that we're not seeing any glyphs, as we asked for textequiv_level == "word"
+    glyphs = tree.xpath("//pc:Glyph", namespaces=NSMAP)
+    assert len(glyphs) == 0
+
+
+def test_glyphs(workspace):
+    CalamariRecognize(
+        workspace,
+        input_file_grp="OCR-D-GT-SEG-LINE",
+        output_file_grp="OCR-D-OCR-CALAMARI",
+        parameter={
+            "checkpoint": CHECKPOINT,
+            "textequiv_level": "glyph",   # Note that we're going down to glyph level here
+        }
+    ).process()
+    workspace.save_mets()
+
+    page1 = os.path.join(workspace.directory, "OCR-D-OCR-CALAMARI/OCR-D-OCR-CALAMARI_0001.xml")
+    assert os.path.exists(page1)
+    tree = etree.parse(page1)
+
+    # The result should contain a lot of glyphs
+    glyphs = tree.xpath("//pc:Glyph", namespaces=NSMAP)
+    assert len(glyphs) >= 100
+
 
 # vim:tw=120: