Pull request #29: RED-5107 robustify image service alternative

Merge in RR/image-prediction from RED-5107-robustify-image-service-alternative to release/1.2.x

Squashed commit of the following:

commit 1a8fbeebd3c05f25d69210e53bf6dce67bc2342f
Merge: 00ac0d6 c03913e
Author: Matthias Bisping <matthias.bisping@axbit.com>
Date:   Tue Aug 30 16:19:16 2022 +0200

    Merge branch 'release/1.2.x' into RED-5107-robustify-image-service-alternative

commit 00ac0d61abdd97eb7c2576d2db9e6859b91c9c41
Author: Matthias Bisping <matthias.bisping@axbit.com>
Date:   Tue Aug 30 16:03:41 2022 +0200

    applied black

commit 983265f4355253a3a371747b04b1926ff3578fef
Author: Matthias Bisping <matthias.bisping@axbit.com>
Date:   Tue Aug 30 15:59:11 2022 +0200

    Added image validation after image extraction to parsable-pdf image extractor. Invalid images are dropped, hence these images will appear as skipped for the service caller.

image_prediction/image_extractor/extractors/parsable.py

@@ -1,14 +1,17 @@
 import atexit
 import io
+import json
+import traceback
 from functools import partial, lru_cache
 from itertools import chain, starmap, filterfalse
-from operator import itemgetter
+from operator import itemgetter, truth
 from typing import List
 
 import fitz
 from PIL import Image
 from funcy import rcompose, merge, pluck, curry, compose
 
+from image_prediction.formatter.formatters.enum import EnumFormatter
 from image_prediction.image_extractor.extractor import ImageExtractor, ImageMetadataPair
 from image_prediction.info import Info
 from image_prediction.stitching.stitching import stitch_pairs
@@ -46,10 +49,28 @@ class ParsablePDFImageExtractor(ImageExtractor):
         clear_caches()
 
         image_metadata_pairs = starmap(ImageMetadataPair, filter(all, zip(images, metadata)))
+        #  TODO: In the future, consider to introduce an image validator as a pipeline component rather than doing the
+        #   validation here. Invalid images can then be split into a different stream and joined with the intact images
+        #   again for the formatting step.
+        image_metadata_pairs = self.__filter_valid_images(image_metadata_pairs)
         image_metadata_pairs = stitch_pairs(list(image_metadata_pairs), tolerance=self.tolerance)
 
         yield from image_metadata_pairs
 
+    @staticmethod
+    def __filter_valid_images(image_metadata_pairs):
+        def validate(image: Image.Image, metadata: dict):
+            try:
+                # TODO: stand-in heuristic for testing if image is valid => find cleaner solution
+                image.resize((100, 100)).convert("RGB")
+                return ImageMetadataPair(image, metadata)
+            except (OSError, Exception) as err:
+                metadata = json.dumps(EnumFormatter()(metadata), indent=2)
+                logger.warning(f"Invalid image encountered. Image metadata:\n{metadata}\n\n{traceback.format_exc()}")
+                return None
+
+        return filter(truth, starmap(validate, image_metadata_pairs))
+
 
 def extract_pages(doc, page_range):
     page_range = range(page_range.start + 1, page_range.stop + 1)

src/serve.py

@@ -10,7 +10,6 @@ from image_prediction.utils.banner import show_banner
 
 
 def main():
-
     def predict(pdf):
         # Keras service_estimator.predict stalls when service_estimator was loaded in different process;
         # therefore, we re-load the model (part of the pipeline) every time we process a new document.

test/conftest.py

@@ -17,7 +17,7 @@ pytest_plugins = [
     "test.fixtures.parameters",
     "test.fixtures.pdf",
     "test.fixtures.target",
-    "test.unit_tests.image_stitching_test"
+    "test.unit_tests.image_stitching_test",
 ]
 
 

test/exploration_tests/funcy_test.py

@@ -3,7 +3,7 @@ from funcy import rcompose, chunks
 
 
 def test_rcompose():
-    f = rcompose(lambda x: x ** 2, str, lambda x: x * 2)
+    f = rcompose(lambda x: x**2, str, lambda x: x * 2)
     assert f(3) == "99"