feat: allow computing page images on-demand and cache them

Signed-off-by: Michele Dolfi <dol@zurich.ibm.com>

docling/backend/docling_parse_backend.py

@@ -80,7 +80,9 @@ class DoclingParsePageBackend(PdfPageBackend):
             cell_counter += 1
 
         def draw_clusters_and_cells():
-            image = self.get_page_image()
+            image = (
+                self.get_page_image()
+            )  # make new image to avoid drawing on the saved ones
             draw = ImageDraw.Draw(image)
             for c in cells:
                 x0, y0, x1, y1 = c.bbox.as_tuple()

docling/backend/pypdfium2_backend.py

@@ -134,7 +134,9 @@ class PyPdfiumPageBackend(PdfPageBackend):
             return merged_cells
 
         def draw_clusters_and_cells():
-            image = self.get_page_image()
+            image = (
+                self.get_page_image()
+            )  # make new image to avoid drawing on the saved ones
             draw = ImageDraw.Draw(image)
             for c in cells:
                 x0, y0, x1, y1 = c.bbox.as_tuple()

docling/datamodel/base_models.py

@@ -234,14 +234,29 @@ class Page(BaseModel):
     model_config = ConfigDict(arbitrary_types_allowed=True)
 
     page_no: int
-    page_hash: str = None
+    page_hash: Optional[str] = None
-    size: PageSize = None
+    size: Optional[PageSize] = None
-    image: Image = None
     cells: List[Cell] = None
     predictions: PagePredictions = PagePredictions()
-    assembled: AssembledUnit = None
+    assembled: Optional[AssembledUnit] = None
 
-    _backend: PdfPageBackend = None  # Internal PDF backend
+    _backend: Optional[PdfPageBackend] = (
+        None  # Internal PDF backend. By default it is cleared during assembling.
+    )
+    _image_cache: Dict[float, Image] = (
+        {}
+    )  # Cache of images in different scales. By default it is cleared during assembling.
+
+    def get_image(self, scale: float = 1.0) -> Optional[Image]:
+        if self._backend is None:
+            return self._image_cache.get(scale, None)
+        if not scale in self._image_cache:
+            self._image_cache[scale] = self._backend.get_page_image(scale=scale)
+        return self._image_cache[scale]
+
+    @property
+    def image(self) -> Optional[Image]:
+        return self.get_image()
 
 
 class DocumentStream(BaseModel):

docling/document_converter.py

@@ -189,9 +189,7 @@ class DocumentConverter:
 
                     # Remove page images (can be disabled)
                     if not self.assemble_options.keep_page_images:
-                        assembled_page.image = (
+                        assembled_page._image_cache = {}
-                            None  # Comment this if you want to visualize page images
-                        )
 
                     # Unload backend
                     assembled_page._backend.unload()
@@ -231,7 +229,7 @@ class DocumentConverter:
 
     # Generate the page image and store it in the page object
     def populate_page_images(self, doc: InputDocument, page: Page) -> Page:
-        page.image = page._backend.get_page_image()
+        page.get_image()  # this will trigger storing the image in the internal cache
 
         return page
 
@@ -247,7 +245,7 @@ class DocumentConverter:
                 draw.rectangle([(x0, y0), (x1, y1)], outline="red")
             image.show()
 
-        # draw_text_boxes(page.image, cells)
+        # draw_text_boxes(page.get_image(scale=1.0), cells)
 
         return page
 

docling/models/easyocr_model.py

@@ -30,7 +30,7 @@ class EasyOcrModel:
 
         for page in page_batch:
             # rects = page._fpage.
-            high_res_image = page._backend.get_page_image(scale=self.scale)
+            high_res_image = page.get_image(scale=self.scale)
             im = numpy.array(high_res_image)
             result = self.reader.readtext(im)
 

docling/models/layout_model.py

@@ -267,7 +267,9 @@ class LayoutModel:
     def __call__(self, page_batch: Iterable[Page]) -> Iterable[Page]:
         for page in page_batch:
             clusters = []
-            for ix, pred_item in enumerate(self.layout_predictor.predict(page.image)):
+            for ix, pred_item in enumerate(
+                self.layout_predictor.predict(page.get_image(scale=1.0))
+            ):
                 cluster = Cluster(
                     id=ix,
                     label=pred_item["label"],

docling/models/table_structure_model.py

@@ -34,7 +34,9 @@ class TableStructureModel:
             self.scale = 2.0  # Scale up table input images to 144 dpi
 
     def draw_table_and_cells(self, page: Page, tbl_list: List[TableElement]):
-        image = page._backend.get_page_image()
+        image = (
+            page._backend.get_page_image()
+        )  # make new image to avoid drawing on the saved ones
         draw = ImageDraw.Draw(image)
 
         for table_element in tbl_list:
@@ -94,13 +96,7 @@ class TableStructureModel:
                 "width": page.size.width * self.scale,
                 "height": page.size.height * self.scale,
             }
-            # add image to page input.
+            page_input["image"] = numpy.asarray(page.get_image(scale=self.scale))
-            if self.scale == 1.0:
-                page_input["image"] = numpy.asarray(page.image)
-            else:  # render new page image on the fly at desired scale
-                page_input["image"] = numpy.asarray(
-                    page._backend.get_page_image(scale=self.scale)
-                )
 
             table_clusters, table_bboxes = zip(*in_tables)