feat(ocr): add image preprocessing module with CLAHE + grayscale + blur

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

ocr-service/preprocessing.py

@@ -0,0 +1,49 @@
+"""Image preprocessing pipeline for aged document OCR quality improvement."""
+
+import logging
+import os
+
+import cv2
+import numpy as np
+from PIL import Image
+
+logger = logging.getLogger(__name__)
+
+CLAHE_CLIP_LIMIT = float(os.environ.get("OCR_CLAHE_CLIP_LIMIT", "2.0"))
+CLAHE_TILE_SIZE = int(os.environ.get("OCR_CLAHE_TILE_SIZE", "8"))
+
+
+def preprocess_page(image: Image.Image) -> Image.Image:
+    """Apply CLAHE + grayscale + Gaussian blur to improve OCR quality on aged documents.
+
+    Falls back silently to the original image if any step fails.
+    """
+    try:
+        img_array = np.array(image)
+        lab = cv2.cvtColor(img_array, cv2.COLOR_RGB2LAB)
+        del img_array
+
+        l_channel = lab[:, :, 0].copy()
+        del lab
+
+        clahe = cv2.createCLAHE(
+            clipLimit=CLAHE_CLIP_LIMIT,
+            tileGridSize=(CLAHE_TILE_SIZE, CLAHE_TILE_SIZE),
+        )
+        l_clahe = clahe.apply(l_channel)
+        del l_channel
+
+        blurred = cv2.GaussianBlur(l_clahe, (3, 3), 0)
+        del l_clahe
+
+        result = Image.fromarray(blurred, mode="L")
+        del blurred
+
+        return result
+    except Exception as e:
+        logger.warning(
+            "preprocess_page failed (falling back to original): %s: %s",
+            type(e).__name__,
+            e,
+        )
+        return image

ocr-service/test_preprocessing.py

@@ -0,0 +1,64 @@
+"""Tests for the image preprocessing pipeline."""
+
+import numpy as np
+import pytest
+from PIL import Image
+from unittest.mock import patch
+
+
+def _make_yellowed_image(width=100, height=100):
+    """Dark, faded yellowed page: L values in a narrow low range with spatial noise.
+
+    Very dark (R≈30, G≈20, B≈10) → L_cv ≈ 80-100 in OpenCV uint8 LAB space.
+    The per-pixel noise gives each CLAHE tile a non-trivial histogram to equalize,
+    which stretches the narrow dark range toward [0-255] and reliably increases mean L.
+    """
+    rng = np.random.default_rng(42)
+    arr = np.zeros((height, width, 3), dtype=np.uint8)
+    arr[:, :, 0] = np.clip(30 + rng.integers(-8, 9, (height, width)), 0, 255)
+    arr[:, :, 1] = np.clip(20 + rng.integers(-5, 6, (height, width)), 0, 255)
+    arr[:, :, 2] = np.clip(10 + rng.integers(-3, 4, (height, width)), 0, 255)
+    return Image.fromarray(arr.astype(np.uint8), mode="RGB")
+
+
+class TestPreprocessPage:
+    def test_output_has_same_dimensions_as_input(self):
+        from preprocessing import preprocess_page
+
+        img = Image.new("RGB", (150, 200))
+        result = preprocess_page(img)
+
+        assert result.size == img.size
+
+    def test_l_channel_mean_increases_on_yellowed_image(self):
+        """CLAHE equalizes the dark narrow-range histogram toward [0-255], raising mean L."""
+        from preprocessing import preprocess_page
+        import cv2
+
+        img = _make_yellowed_image()
+
+        arr_before = np.array(img)
+        lab_before = cv2.cvtColor(arr_before, cv2.COLOR_RGB2LAB)
+        l_mean_before = float(lab_before[:, :, 0].mean())
+
+        result = preprocess_page(img)
+
+        # Output is grayscale (mode "L"); its values ARE the CLAHE-enhanced L channel
+        l_mean_after = float(np.array(result).mean())
+
+        assert l_mean_after > l_mean_before
+
+    def test_falls_back_to_pixel_identical_original_on_cv2_error(self):
+        """When cv2 raises, preprocess_page must return the unmodified original image."""
+        from preprocessing import preprocess_page
+
+        img = Image.new("RGB", (80, 60), color=(123, 45, 67))
+        original_pixels = list(img.getdata())
+
+        with patch("preprocessing.cv2") as mock_cv2:
+            mock_cv2.cvtColor.side_effect = RuntimeError("cv2 exploded")
+
+            result = preprocess_page(img)
+
+        result_pixels = list(result.getdata())
+        assert result_pixels == original_pixels