feat: ocr when too many chars unrecognized (#8)

Certain PDF documents have text blocks within their layout that contain text, but the text contains / consists of unrecognized characters, represented in the text as (cid:n) where n is an integer.

Current logic accepts whatever text is present, only applying OCR in the case the layout text is None. This PR changes the logic to also apply OCR in the case that > 50% of the characters are unrecognized.

Author: qued

CHANGELOG.md

@@ -1,3 +1,7 @@
+## 0.2.2-dev0
+
+* Add logic to use OCR when layout text is full of unknown characters
+
 ## 0.2.1
 
 * Refactor to facilitate local inference

test_unstructured_inference/inference/test_layout.py

@@ -176,3 +176,39 @@ def test_process_file_with_model(monkeypatch, mock_page_layout, model_name):
 def test_process_file_with_model_raises_on_invalid_model_name():
     with pytest.raises(models.UnknownModelException):
         layout.process_file_with_model("", model_name="fake")
+
+
+class MockPageLayout(layout.PageLayout):
+    def __init__(self, ocr_text):
+        self.ocr_text = ocr_text
+
+    def ocr(self, text_block):
+        return self.ocr_text
+
+
+class MockTextBlock(lp.TextBlock):
+    def __init__(self, text):
+        self.text = text
+
+
+def test_interpret_text_block_use_ocr_when_text_symbols_cid():
+    fake_text = "(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)"
+    fake_ocr = "ocrme"
+    fake_text_block = MockTextBlock(fake_text)
+    assert MockPageLayout(fake_ocr).interpret_text_block(fake_text_block) == fake_ocr
+
+
+@pytest.mark.parametrize(
+    "text, expected",
+    [("base", 0.0), ("", 0.0), ("(cid:2)", 1.0), ("(cid:1)a", 0.5), ("c(cid:1)ab", 0.25)],
+)
+def test_cid_ratio(text, expected):
+    assert layout.cid_ratio(text) == expected
+
+
+@pytest.mark.parametrize(
+    "text, expected",
+    [("base", False), ("(cid:2)", True), ("(cid:1234567890)", True), ("jkl;(cid:12)asdf", True)],
+)
+def test_is_cid_present(text, expected):
+    assert layout.is_cid_present(text) == expected

unstructured_inference/__version__.py

@@ -1 +1 @@
-__version__ = "0.2.1"  # pragma: no cover
+__version__ = "0.2.2-dev0"  # pragma: no cover

unstructured_inference/inference/layout.py

@@ -1,5 +1,6 @@
 from __future__ import annotations
 from dataclasses import dataclass
+import re
 import tempfile
 from typing import List, Optional, Tuple, Union, BinaryIO
 
@@ -109,10 +110,7 @@ def get_elements(self, inplace=True) -> Optional[List[LayoutElement]]:
             text_blocks = self.layout.filter_by(item, center=True)
             text = str()
             for text_block in text_blocks:
-                # NOTE(robinson) - If the text attribute is None, that means the PDF isn't
-                # already OCR'd and we have to send the snippet out for OCRing.
-                if text_block.text is None:
-                    text_block.text = self.ocr(text_block)
+                text_block.text = self.interpret_text_block(text_block)
             text = " ".join([x for x in text_blocks.get_texts() if x])
 
             elements.append(
@@ -124,6 +122,16 @@ def get_elements(self, inplace=True) -> Optional[List[LayoutElement]]:
             return None
         return elements
 
+    def interpret_text_block(self, text_block: lp.TextBlock) -> str:
+        """Interprets the text in a TextBlock."""
+        # NOTE(robinson) - If the text attribute is None, that means the PDF isn't
+        # already OCR'd and we have to send the snippet out for OCRing.
+        if (text_block.text is None) or cid_ratio(text_block.text) > 0.5:
+            out_text = self.ocr(text_block)
+        else:
+            out_text = text_block.text
+        return out_text
+
     def ocr(self, text_block: lp.TextBlock) -> str:
         """Runs a cropped text block image through and OCR agent."""
         logger.debug("Running OCR on text block ...")
@@ -156,3 +164,19 @@ def process_file_with_model(filename: str, model_name: str) -> DocumentLayout:
     model = None if model_name is None else get_model(model_name)
     layout = DocumentLayout.from_file(filename, model=model)
     return layout
+
+
+def cid_ratio(text: str) -> float:
+    """Gets ratio of unknown 'cid' characters extracted from text to all characters."""
+    if not is_cid_present(text):
+        return 0.0
+    cid_pattern = r"\(cid\:(\d+)\)"
+    unmatched, n_cid = re.subn(cid_pattern, "", text)
+    total = n_cid + len(unmatched)
+    return n_cid / total
+
+
+def is_cid_present(text: str) -> bool:
+    if len(text) < len("(cid:x)"):
+        return False
+    return text.find("(cid:") != -1