feat(OCR): Introduce the OcrOptions.force_full_page_ocr parameter that forces a full page OCR scanning (#290)

- When the OCR is forced, any existing PDF cells are rejected.
- Introduce the force-ocr cmd parameter in docling CLI.
- Update unit tests.
- Add the full_page_ocr.py example in mkdocs.

Signed-off-by: Nikos Livathinos <nli@zurich.ibm.com>

docling/cli/main.py

@@ -153,6 +153,13 @@ def convert(
             ..., help="If enabled, the bitmap content will be processed using OCR."
         ),
     ] = True,
+    force_ocr: Annotated[
+        bool,
+        typer.Option(
+            ...,
+            help="Replace any existing text with OCR generated text over the full content.",
+        ),
+    ] = False,
     ocr_engine: Annotated[
         OcrEngine, typer.Option(..., help="The OCR engine to use.")
     ] = OcrEngine.EASYOCR,
@@ -219,11 +226,11 @@ def convert(
 
     match ocr_engine:
         case OcrEngine.EASYOCR:
-            ocr_options: OcrOptions = EasyOcrOptions()
+            ocr_options: OcrOptions = EasyOcrOptions(force_full_page_ocr=force_ocr)
         case OcrEngine.TESSERACT_CLI:
-            ocr_options = TesseractCliOcrOptions()
+            ocr_options = TesseractCliOcrOptions(force_full_page_ocr=force_ocr)
         case OcrEngine.TESSERACT:
-            ocr_options = TesseractOcrOptions()
+            ocr_options = TesseractOcrOptions(force_full_page_ocr=force_ocr)
         case _:
             raise RuntimeError(f"Unexpected OCR engine type {ocr_engine}")
 

docling/datamodel/pipeline_options.py

@@ -22,6 +22,7 @@ class TableStructureOptions(BaseModel):
 
 class OcrOptions(BaseModel):
     kind: str
+    force_full_page_ocr: bool = False  # If enabled a full page OCR is always applied
     bitmap_area_threshold: float = (
         0.05  # percentage of the area for a bitmap to processed with OCR
     )

docling/models/base_ocr_model.py

@@ -10,7 +10,7 @@ from PIL import Image, ImageDraw
 from rtree import index
 from scipy.ndimage import find_objects, label
 
-from docling.datamodel.base_models import OcrCell, Page
+from docling.datamodel.base_models import Cell, OcrCell, Page
 from docling.datamodel.document import ConversionResult
 from docling.datamodel.pipeline_options import OcrOptions
 from docling.datamodel.settings import settings
@@ -73,7 +73,9 @@ class BaseOcrModel(BasePageModel):
         coverage, ocr_rects = find_ocr_rects(page.size, bitmap_rects)
 
         # return full-page rectangle if sufficiently covered with bitmaps
-        if coverage > max(BITMAP_COVERAGE_TRESHOLD, self.options.bitmap_area_threshold):
+        if self.options.force_full_page_ocr or coverage > max(
+            BITMAP_COVERAGE_TRESHOLD, self.options.bitmap_area_threshold
+        ):
             return [
                 BoundingBox(
                     l=0,
@@ -96,7 +98,7 @@ class BaseOcrModel(BasePageModel):
             return ocr_rects
 
     # Filters OCR cells by dropping any OCR cell that intersects with an existing programmatic cell.
-    def filter_ocr_cells(self, ocr_cells, programmatic_cells):
+    def _filter_ocr_cells(self, ocr_cells, programmatic_cells):
         # Create R-tree index for programmatic cells
         p = index.Property()
         p.dimension = 2
@@ -117,6 +119,23 @@ class BaseOcrModel(BasePageModel):
         ]
         return filtered_ocr_cells
 
+    def post_process_cells(self, ocr_cells, programmatic_cells):
+        r"""
+        Post-process the ocr and programmatic cells and return the final list of of cells
+        """
+        if self.options.force_full_page_ocr:
+            # If a full page OCR is forced, use only the OCR cells
+            cells = [
+                Cell(id=c_ocr.id, text=c_ocr.text, bbox=c_ocr.bbox)
+                for c_ocr in ocr_cells
+            ]
+            return cells
+
+        ## Remove OCR cells which overlap with programmatic cells.
+        filtered_ocr_cells = self._filter_ocr_cells(ocr_cells, programmatic_cells)
+        programmatic_cells.extend(filtered_ocr_cells)
+        return programmatic_cells
+
     def draw_ocr_rects_and_cells(self, conv_res, page, ocr_rects, show: bool = False):
         image = copy.deepcopy(page.image)
         draw = ImageDraw.Draw(image, "RGBA")

docling/models/easyocr_model.py

@@ -5,7 +5,7 @@ import numpy
 import torch
 from docling_core.types.doc import BoundingBox, CoordOrigin
 
-from docling.datamodel.base_models import OcrCell, Page
+from docling.datamodel.base_models import Cell, OcrCell, Page
 from docling.datamodel.document import ConversionResult
 from docling.datamodel.pipeline_options import EasyOcrOptions
 from docling.datamodel.settings import settings
@@ -88,12 +88,8 @@ class EasyOcrModel(BaseOcrModel):
                         ]
                         all_ocr_cells.extend(cells)
 
-                    ## Remove OCR cells which overlap with programmatic cells.
+                    # Post-process the cells
-                    filtered_ocr_cells = self.filter_ocr_cells(
+                    page.cells = self.post_process_cells(all_ocr_cells, page.cells)
-                        all_ocr_cells, page.cells
-                    )
-
-                    page.cells.extend(filtered_ocr_cells)
 
                 # DEBUG code:
                 if settings.debug.visualize_ocr:

docling/models/tesseract_ocr_cli_model.py

@@ -7,7 +7,7 @@ from typing import Iterable, Optional, Tuple
 import pandas as pd
 from docling_core.types.doc import BoundingBox, CoordOrigin
 
-from docling.datamodel.base_models import OcrCell, Page
+from docling.datamodel.base_models import Cell, OcrCell, Page
 from docling.datamodel.document import ConversionResult
 from docling.datamodel.pipeline_options import TesseractCliOcrOptions
 from docling.datamodel.settings import settings
@@ -170,12 +170,8 @@ class TesseractOcrCliModel(BaseOcrModel):
                             )
                             all_ocr_cells.append(cell)
 
-                    ## Remove OCR cells which overlap with programmatic cells.
+                    # Post-process the cells
-                    filtered_ocr_cells = self.filter_ocr_cells(
+                    page.cells = self.post_process_cells(all_ocr_cells, page.cells)
-                        all_ocr_cells, page.cells
-                    )
-
-                    page.cells.extend(filtered_ocr_cells)
 
                 # DEBUG code:
                 if settings.debug.visualize_ocr:

docling/models/tesseract_ocr_model.py

@@ -3,7 +3,7 @@ from typing import Iterable
 
 from docling_core.types.doc import BoundingBox, CoordOrigin
 
-from docling.datamodel.base_models import OcrCell, Page
+from docling.datamodel.base_models import Cell, OcrCell, Page
 from docling.datamodel.document import ConversionResult
 from docling.datamodel.pipeline_options import TesseractOcrOptions
 from docling.datamodel.settings import settings
@@ -140,12 +140,8 @@ class TesseractOcrModel(BaseOcrModel):
                         # del high_res_image
                         all_ocr_cells.extend(cells)
 
-                    ## Remove OCR cells which overlap with programmatic cells.
+                    # Post-process the cells
-                    filtered_ocr_cells = self.filter_ocr_cells(
+                    page.cells = self.post_process_cells(all_ocr_cells, page.cells)
-                        all_ocr_cells, page.cells
-                    )
-
-                    page.cells.extend(filtered_ocr_cells)
 
                 # DEBUG code:
                 if settings.debug.visualize_ocr:

docs/examples/full_page_ocr.py

@@ -0,0 +1,42 @@
+from pathlib import Path
+
+from docling.backend.docling_parse_backend import DoclingParseDocumentBackend
+from docling.datamodel.base_models import InputFormat
+from docling.datamodel.pipeline_options import (
+    EasyOcrOptions,
+    PdfPipelineOptions,
+    TesseractCliOcrOptions,
+    TesseractOcrOptions,
+)
+from docling.document_converter import DocumentConverter, PdfFormatOption
+
+
+def main():
+    input_doc = Path("./tests/data/2206.01062.pdf")
+
+    pipeline_options = PdfPipelineOptions()
+    pipeline_options.do_ocr = True
+    pipeline_options.do_table_structure = True
+    pipeline_options.table_structure_options.do_cell_matching = True
+
+    # Any of the OCR options can be used:EasyOcrOptions, TesseractOcrOptions, TesseractCliOcrOptions
+    # ocr_options = EasyOcrOptions(force_full_page_ocr=True)
+    # ocr_options = TesseractOcrOptions(force_full_page_ocr=True)
+    ocr_options = TesseractCliOcrOptions(force_full_page_ocr=True)
+    pipeline_options.ocr_options = ocr_options
+
+    converter = DocumentConverter(
+        format_options={
+            InputFormat.PDF: PdfFormatOption(
+                pipeline_options=pipeline_options,
+            )
+        }
+    )
+
+    doc = converter.convert(input_doc).document
+    md = doc.export_to_markdown()
+    print(md)
+
+
+if __name__ == "__main__":
+    main()

mkdocs.yml

@@ -71,6 +71,7 @@ nav:
       - "Figure enrichment": examples/develop_picture_enrichment.py
       - "Table export": examples/export_tables.py
       - "Multimodal export": examples/export_multimodal.py
+      - "Force full page OCR": examples/full_page_ocr.py
     - RAG / QA:
       - "RAG with LlamaIndex 🦙": examples/rag_llamaindex.ipynb
       - "RAG with LangChain 🦜🔗": examples/rag_langchain.ipynb

tests/test_e2e_ocr_conversion.py

@@ -15,34 +15,8 @@ from docling.document_converter import DocumentConverter, PdfFormatOption
 
 from .verify_utils import verify_conversion_result_v1, verify_conversion_result_v2
 
-GENERATE = False
+GENERATE_V1 = False
-
+GENERATE_V2 = False
-
-# Debug
-def save_output(pdf_path: Path, doc_result: ConversionResult, engine: str):
-    r""" """
-    import json
-    import os
-
-    parent = pdf_path.parent
-    eng = "" if engine is None else f".{engine}"
-
-    dict_fn = os.path.join(parent, f"{pdf_path.stem}{eng}.json")
-    with open(dict_fn, "w") as fd:
-        json.dump(doc_result.legacy_document.export_to_dict(), fd)
-
-    pages_fn = os.path.join(parent, f"{pdf_path.stem}{eng}.pages.json")
-    pages = [p.model_dump() for p in doc_result.pages]
-    with open(pages_fn, "w") as fd:
-        json.dump(pages, fd)
-
-    doctags_fn = os.path.join(parent, f"{pdf_path.stem}{eng}.doctags.txt")
-    with open(doctags_fn, "w") as fd:
-        fd.write(doc_result.legacy_document.export_to_doctags())
-
-    md_fn = os.path.join(parent, f"{pdf_path.stem}{eng}.md")
-    with open(md_fn, "w") as fd:
-        fd.write(doc_result.legacy_document.export_to_markdown())
 
 
 def get_pdf_paths():
@@ -74,13 +48,15 @@ def get_converter(ocr_options: OcrOptions):
 
 
 def test_e2e_conversions():
-
     pdf_paths = get_pdf_paths()
 
     engines: List[OcrOptions] = [
         EasyOcrOptions(),
         TesseractOcrOptions(),
         TesseractCliOcrOptions(),
+        EasyOcrOptions(force_full_page_ocr=True),
+        TesseractOcrOptions(force_full_page_ocr=True),
+        TesseractCliOcrOptions(force_full_page_ocr=True),
     ]
 
     for ocr_options in engines:
@@ -91,20 +67,16 @@ def test_e2e_conversions():
 
             doc_result: ConversionResult = converter.convert(pdf_path)
 
-            # Save conversions
-            # save_output(pdf_path, doc_result, None)
-
-            # Debug
             verify_conversion_result_v1(
                 input_path=pdf_path,
                 doc_result=doc_result,
-                generate=GENERATE,
+                generate=GENERATE_V1,
                 fuzzy=True,
             )
 
             verify_conversion_result_v2(
                 input_path=pdf_path,
                 doc_result=doc_result,
-                generate=GENERATE,
+                generate=GENERATE_V2,
                 fuzzy=True,
             )

tests/verify_utils.py

@@ -256,15 +256,19 @@ def verify_conversion_result_v1(
     dt_path = gt_subpath.with_suffix(f"{engine_suffix}.doctags.txt")
 
     if generate:  # only used when re-generating truth
+        pages_path.parent.mkdir(parents=True, exist_ok=True)
         with open(pages_path, "w") as fw:
             fw.write(json.dumps(doc_pred_pages, default=pydantic_encoder))
 
+        json_path.parent.mkdir(parents=True, exist_ok=True)
         with open(json_path, "w") as fw:
             fw.write(json.dumps(doc_pred, default=pydantic_encoder))
 
+        md_path.parent.mkdir(parents=True, exist_ok=True)
         with open(md_path, "w") as fw:
             fw.write(doc_pred_md)
 
+        dt_path.parent.mkdir(parents=True, exist_ok=True)
         with open(dt_path, "w") as fw:
             fw.write(doc_pred_dt)
     else:  # default branch in test
@@ -328,15 +332,19 @@ def verify_conversion_result_v2(
     dt_path = gt_subpath.with_suffix(f"{engine_suffix}.doctags.txt")
 
     if generate:  # only used when re-generating truth
+        pages_path.parent.mkdir(parents=True, exist_ok=True)
         with open(pages_path, "w") as fw:
             fw.write(json.dumps(doc_pred_pages, default=pydantic_encoder))
 
+        json_path.parent.mkdir(parents=True, exist_ok=True)
         with open(json_path, "w") as fw:
             fw.write(json.dumps(doc_pred, default=pydantic_encoder))
 
+        md_path.parent.mkdir(parents=True, exist_ok=True)
         with open(md_path, "w") as fw:
             fw.write(doc_pred_md)
 
+        dt_path.parent.mkdir(parents=True, exist_ok=True)
         with open(dt_path, "w") as fw:
             fw.write(doc_pred_dt)
     else:  # default branch in test