Use PdfText for hi_res strategy

unstructured/partition/pdf.py

@@ -72,6 +72,7 @@ from unstructured.partition.pdf_image.pdfminer_processing import (
     clean_pdfminer_duplicate_image_elements,
     clean_pdfminer_inner_elements,
     merge_inferred_with_extracted_layout,
+    _extract_text_pdftext
 )
 from unstructured.partition.pdf_image.pdfminer_utils import (
     open_pdfminer_pages_generator,
@@ -686,17 +687,6 @@ def pdfminer_interpreter_init_resources(wrapped, instance, args, kwargs):
 
     return wrapped(resources)
 
-# Simple implementation, combines blocks into one singe text element
-# Each line ends with \n so it's possible to easily split them if needed
-def _extract_text_pdftext(lines: list[dict[str, Any]]) -> str:
-    text = ""
-    
-    for line in lines:
-        for span in line["spans"]:
-            text += span["text"]
-            
-    return text
-
 # This function is not meant to be used right away
 # Needs better implementation but the point is that
 # it's possible to extracts URL annotations using pydfium2
@@ -741,6 +731,7 @@ def _process_pdfminer_pages(
 
     elements: list[Element] = []
     
+    # Open the PDF file using pypdfium2
     pdf = pdfium.PdfDocument(fp)
 
     for page_number, page in enumerate(
@@ -761,7 +752,10 @@ def _process_pdfminer_pages(
         #     annotation_list = get_uris(page.annots, height, coordinate_system, page_number)
 
         for obj in page["blocks"]:
+            # Not sure if rect_to_bbox function shouldn't be used here
+            # x1, y1, x2, y2 = rect_to_bbox(obj.bbox, height)
             x1, y1, x2, y2 = obj['bbox']
+            
             # bbox = (x1, y1, x2, y2)
 
             # urls_metadata: list[dict[str, Any]] = []

unstructured/partition/pdf_image/pdfminer_processing.py

@@ -1,4 +1,4 @@
-from typing import TYPE_CHECKING, BinaryIO, List, Optional, Union, cast
+from typing import TYPE_CHECKING, BinaryIO, List, Optional, Union, cast, Any
 
 from pdfminer.utils import open_filename
 
@@ -17,12 +17,16 @@ if TYPE_CHECKING:
     from unstructured_inference.inference.elements import TextRegion
     from unstructured_inference.inference.layout import DocumentLayout
 
+from pdftext.extraction import dictionary_output
+import pypdfium2 as pdfium
+
 
 def process_file_with_pdfminer(
     filename: str = "",
     dpi: int = 200,
 ) -> List[List["TextRegion"]]:
-    with open_filename(filename, "rb") as fp:
+    # Only reason to change this is to not use PDFminer functions
+    with open(filename, "rb") as fp:
         fp = cast(BinaryIO, fp)
         extracted_layout = process_data_with_pdfminer(
             file=fp,
@@ -31,6 +35,18 @@ def process_file_with_pdfminer(
         return extracted_layout
 
 
+# Simple implementation, combines blocks into one singe text element
+# Each line ends with \n so it's possible to easily split them if needed
+def _extract_text_pdftext(lines: list[dict[str, Any]]) -> str:
+    text = ""
+
+    for line in lines:
+        for span in line["spans"]:
+            text += span["text"]
+
+    return text
+
+
 @requires_dependencies("unstructured_inference")
 def process_data_with_pdfminer(
     file: Optional[Union[bytes, BinaryIO]] = None,
@@ -46,38 +62,53 @@ def process_data_with_pdfminer(
     from unstructured_inference.inference.ordering import order_layout
 
     layouts = []
+
+    # Open the PDF file using pypdfium2
+    pdf = pdfium.PdfDocument(file)
     # Coefficient to rescale bounding box to be compatible with images
     coef = dpi / 72
-    for page, page_layout in open_pdfminer_pages_generator(file):
-        height = page_layout.height
+
+    for page_index, page in enumerate(dictionary_output(pdf, sort=False)):
+        height = page["height"]
 
         layout: List["TextRegion"] = []
-        for obj in page_layout:
-            x1, y1, x2, y2 = rect_to_bbox(obj.bbox, height)
 
-            if hasattr(obj, "get_text"):
-                _text = obj.get_text()
-                element_class = EmbeddedTextRegion  # type: ignore
-            else:
-                embedded_images = get_images_from_pdf_element(obj)
-                if len(embedded_images) > 0:
-                    _text = None
-                    element_class = ImageTextRegion  # type: ignore
-                else:
-                    continue
+        # Since PdfText doesn't contain images we extract text only first
+        for obj in page["blocks"]:
+            # Not sure if rect_to_bbox function shouldn't be used here
+            # x1, y1, x2, y2 = rect_to_bbox(obj.bbox, height)
+            x1, y1, x2, y2 = obj["bbox"]
 
-            text_region = element_class.from_coords(
+            _text = _extract_text_pdftext(obj["lines"])
+
+            text_region = EmbeddedTextRegion.from_coords(
                 x1 * coef,
                 y1 * coef,
                 x2 * coef,
                 y2 * coef,
                 text=_text,
-                source=Source.PDFMINER,
+                source="pdftext",
             )
 
             if text_region.bbox is not None and text_region.bbox.area > 0:
                 layout.append(text_region)
 
+        for obj in page[page_index].get_objects():
+            if isinstance(obj, pdfium.PdfImage) and obj.type == 3:
+                # Not sure if rect_to_bbox function shouldn't be used here
+                # x1, y1, x2, y2 = rect_to_bbox(obj.get_pos(), height)
+                x1, y1, x2, y2 = obj.get_pos()
+                image_region = ImageTextRegion.from_coords(
+                    x1 * coef,
+                    y1 * coef,
+                    x2 * coef,
+                    y2 * coef,
+                    text=None,
+                    source="pdftext",
+                )
+                if image_region.bbox is not None and image_region.bbox.area > 0:
+                    layout.append(image_region)
+
         # NOTE(christine): always do the basic sort first for deterministic order across
         # python versions.
         layout = order_layout(layout)