feat: RED-10765: ignore perceptual hash for image deduplication and prefer to keep the ones with allPassed set to True

feat: RED-10765: filter out classifications for 'duplicate' images present in the document

Closes RED-10765

See merge request redactmanager/image-classification-service!23

config/settings.toml

@@ -4,7 +4,7 @@ level = "INFO"
 [service]
 # Print document processing progress to stdout
 verbose = false
-batch_size = 16
+batch_size = 6
 image_stiching_tolerance = 1  # in pixels
 mlflow_run_id = "fabfb1f192c745369b88cab34471aba7"
 
@@ -36,4 +36,7 @@ max = 10
 [filters.overrides.signature.image_to_page_quotient]
 max = 0.4
 
+[filters.overrides.logo.image_to_page_quotient]
+min = 0.06
+
 

pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "image-classification-service"
-version = "2.16.0"
+version = "2.17.0"
 description = ""
 authors = ["Team Research <research@knecon.com>"]
 readme = "README.md"
@@ -10,8 +10,8 @@ packages = [{ include = "image_prediction", from = "src" }]
 python = ">=3.10,<3.11"
 # FIXME: This should be recent pyinfra, but the recent protobuf packages are not compatible with tensorflow 2.9.0, also
 #  see RED-9948.
-pyinfra = { version = "3.3.5", source = "gitlab-research" }
-kn-utils = { version = ">=0.3.2,<0.4.0", source = "gitlab-research" }
+pyinfra = { version = "3.4.2", source = "gitlab-research" }
+kn-utils = { version = ">=0.4.0", source = "gitlab-research" }
 dvc = "^2.34.0"
 dvc-ssh = "^2.20.0"
 dvc-azure = "^2.21.2"

scripts/run_pipeline.py

@@ -10,6 +10,8 @@ from image_prediction.utils.pdf_annotation import annotate_pdf
 
 logger = get_logger()
 
+logger.setLevel("DEBUG")
+
 
 def parse_args():
     parser = argparse.ArgumentParser()

src/image_prediction/pipeline.py

@@ -1,6 +1,7 @@
 import os
 from functools import lru_cache, partial
 from itertools import chain, tee
+from typing import Iterable, Any
 
 from funcy import rcompose, first, compose, second, chunks, identity, rpartial
 from kn_utils.logging import logger
@@ -54,7 +55,7 @@ class Pipeline:
         join = compose(starlift(lambda prd, rpr, mdt: {"classification": prd, **mdt, "representation": rpr}), star(zip))
 
         #                       />--classify--\
-        # --extract-->--split--+->--encode---->+--join-->reformat
+        # --extract-->--split--+->--encode---->+--join-->reformat-->filter_duplicates
         #                       \>--identity--/
 
         self.pipe = rcompose(
@@ -63,6 +64,7 @@ class Pipeline:
             pairwise_apply(classify, represent, identity),  # ... apply functions to the streams pairwise
             join,  # ... the streams by zipping
             reformat,  # ... the items
+            filter_duplicates,  # ... filter out duplicate images
         )
 
     def __call__(self, pdf: bytes, page_range: range = None):
@@ -72,3 +74,32 @@ class Pipeline:
             unit=" images",
             disable=not self.verbose,
         )
+
+
+def filter_duplicates(metadata: Iterable[dict[str, Any]]) -> Iterable[dict[str, Any]]:
+    """Filter out duplicate images from the `position` (image coordinates) and `page`, preferring the one with
+    `allPassed` set to True.
+    See RED-10765 (RM-241): Removed redactions reappear for why this is necessary.
+    """
+    keep = dict()
+    for image_meta in metadata:
+        key: tuple[int, int, int, int, int] = (
+            image_meta["position"]["x1"],
+            image_meta["position"]["x2"],
+            image_meta["position"]["y1"],
+            image_meta["position"]["y2"],
+            image_meta["position"]["pageNumber"],
+        )
+        if key in keep:
+            logger.warning(
+                f"Duplicate image found: x1={key[0]}, x2={key[1]}, y1={key[2]}, y2={key[3]}, pageNumber={key[4]}"
+            )
+            if image_meta["filters"]["allPassed"]:
+                logger.warning("Setting the image with allPassed flag set to True")
+                keep[key] = image_meta
+            else:
+                logger.warning("Keeping the previous image since the current image has allPassed flag set to False")
+        else:
+            keep[key] = image_meta
+
+    yield from keep.values()

test/data.dvc

@@ -1,5 +1,5 @@
 outs:
-- md5: ab352d3b2c62ce2293cafb57c1b41b01.dir
-  size: 7469082
-  nfiles: 6
+- md5: 08bf8a63f04b3f19f859008556699708.dir
+  size: 7979836
+  nfiles: 7
   path: data

test/regressions_tests/image_deduplication_test.py

@@ -0,0 +1,35 @@
+from pathlib import Path
+
+from image_prediction.config import CONFIG
+from image_prediction.pipeline import load_pipeline
+
+
+def test_all_duplicate_images_are_filtered():
+    """See RED-10765 (RM-241): Removed redactions reappear."""
+    pdf_path = (
+        Path(__file__).parents[1]
+        / "data"
+        / "RED-10765"
+        / "RM-241-461c90d6d6dc0416ad5f0b05feef4dfc.UNTOUCHED_shortened.pdf"
+    )
+    pdf_bytes = pdf_path.read_bytes()
+
+    pipeline = load_pipeline(verbose=True, batch_size=CONFIG.service.batch_size)
+    predictions = list(pipeline(pdf_bytes))
+
+    seen = set()
+    for prediction in predictions:
+        key = (
+            prediction["position"]["x1"],
+            prediction["position"]["x2"],
+            prediction["position"]["y1"],
+            prediction["position"]["y2"],
+            prediction["position"]["pageNumber"],
+        )
+        assert key not in seen, f"Duplicate found: {key}"
+        seen.add(key)
+
+    all_passed = sum(1 for prediction in predictions if prediction["filters"]["allPassed"])
+    assert all_passed == 1, f"Expected 1 image with allPassed flag set to True, but got {all_passed}"
+
+    assert len(predictions) == 177, f"Expected 177 images, but got {len(predictions)}"