Pull request #30: RED-5202 add per class customizable max_rel_image_size configuration.

Merge in RR/image-prediction from RED-5202-add-class-specific-image-size-heuristic to release/1.2.x

* commit 'f48455c496a12d99bb8e9e015d61f661bd94519b':
  RED-5202 add per class customizable max_rel_image_size configuration. If value isn't set or isn't defined at all for a class, the default value is used (change is backwards compatible). Custom value enabled for signatures, set to 0.4 relative image size if not overwritten by ENV (; set to null to use default value or define custom value. Value is 0.4 if ENV is not found.)

config.yaml

@@ -18,6 +18,10 @@ filters:
   image_to_page_quotient: # Image size to page size ratio (ratio of geometric means of areas)
     min: $MIN_REL_IMAGE_SIZE|0.05  # Minimum permissible
     max: $MAX_REL_IMAGE_SIZE|0.75  # Maximum permissible
+    customized:  # Customized settings per class (RED-5202)
+      max:
+        signature: $MAX_REL_SIGNATURE_SIZE|0.4
+
 
   image_width_to_height_quotient:  # Image width to height ratio
     min: $MIN_IMAGE_FORMAT|0.1  # Minimum permissible

image_prediction/config.py

@@ -15,6 +15,12 @@ class DotIndexable:
     def __init__(self, x):
         self.x = x
 
+    def get(self, item, default=None):
+        try:
+            return _get_item_and_maybe_make_dotindexable(self.x, item)
+        except KeyError:
+            return default
+
     def __getattr__(self, item):
         return _get_item_and_maybe_make_dotindexable(self.x, item)
 

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)

image_prediction/transformer/transformers/response.py

@@ -20,6 +20,13 @@ def build_image_info(data: dict) -> dict:
         image_area_sqrt = math.sqrt(abs(x2 - x1) * abs(y2 - y1))
         return image_area_sqrt / page_area_sqrt
 
+    def is_max_image_to_page_quotient_breached(quotient, label):
+        default_max_quotient = CONFIG.filters.image_to_page_quotient.max
+        customized_entries = CONFIG.filters.image_to_page_quotient.customized.max
+        max_quotient = customized_entries.get(label, default_max_quotient)
+        max_quotient = max_quotient if max_quotient else default_max_quotient
+        return bool(quotient > max_quotient)
+
     page_width, page_height, x1, x2, y1, y2, width, height, alpha = itemgetter(
         "page_width", "page_height", "x1", "x2", "y1", "y2", "width", "height", "alpha"
     )(data)
@@ -27,7 +34,9 @@ def build_image_info(data: dict) -> dict:
     quotient = round(compute_geometric_quotient(), 4)
 
     min_image_to_page_quotient_breached = bool(quotient < CONFIG.filters.image_to_page_quotient.min)
-    max_image_to_page_quotient_breached = bool(quotient > CONFIG.filters.image_to_page_quotient.max)
+    max_image_to_page_quotient_breached = is_max_image_to_page_quotient_breached(
+        quotient, data["classification"]["label"]
+    )
     min_image_width_to_height_quotient_breached = bool(
         width / height < CONFIG.filters.image_width_to_height_quotient.min
     )

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"