improve performance

- disable scanned page filter, since dropping these disables the
computation of the images hash and the frontend OCR hint, which are both
wanted
- optimize image extraction by using arrays instead of byte streams for
the conversion to PIL images

image_prediction/image_extractor/extractors/parsable.py

@@ -1,5 +1,4 @@
 import atexit
-import io
 import json
 import traceback
 from _operator import itemgetter
@@ -9,6 +8,7 @@ from operator import itemgetter, truth
 from typing import Iterable, Iterator, List
 
 import fitz
+import numpy as np
 from PIL import Image
 from funcy import merge, pluck, curry, compose, rcompose, remove
 
@@ -111,7 +111,6 @@ def get_metadata_for_images_on_page(page: fitz.Page):
 
 def filter_valid_metadata(metadata):
     yield from compose(
-        filter_out_page_sized_images,
         filter_out_tiny_images,
         filter_invalid_metadata,
     )(metadata)
@@ -136,8 +135,9 @@ def get_image_infos(page: fitz.Page) -> List[dict]:
 
 @lru_cache(maxsize=None)
 def xref_to_image(doc, xref) -> Image:
-    maybe_image = load_image_handle_from_xref(doc, xref)
-    return Image.open(io.BytesIO(maybe_image["image"])) if maybe_image else None
+    pixmap = fitz.Pixmap(doc, xref)
+    array = np.frombuffer(pixmap.samples, dtype=np.uint8).reshape(pixmap.h, pixmap.w)
+    return Image.fromarray(array)
 
 
 def get_image_metadata(image_info):