feat(scan): whiten receipt backgrounds and embed OCR in PDFs

agents/fedeo-device-agent/scripts/scan_background.py

@@ -0,0 +1,74 @@
+"""Conservative paper masking. Keep the canvas and every pixel inside the paper."""
+from pathlib import Path
+
+import cv2
+import numpy as np
+
+
+def whiten_background(image):
+    height, width = image.shape[:2]
+    scale = min(1.0, 1000.0 / max(height, width))
+    small = cv2.resize(image, None, fx=scale, fy=scale, interpolation=cv2.INTER_AREA)
+    gray = cv2.cvtColor(small, cv2.COLOR_BGR2GRAY)
+    border = np.concatenate((gray[0], gray[-1], gray[:, 0], gray[:, -1]))
+    background = float(np.median(border))
+    # White surroundings already need no correction. Never search inside a
+    # white document for an isolated text box or logo to keep instead.
+    if background >= 235 or np.percentile(border, 75) - np.percentile(border, 25) > 55:
+        return image
+
+    threshold, _ = cv2.threshold(gray, 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)
+    mask = np.uint8(gray > max(threshold, background + 20)) * 255
+    mask = cv2.morphologyEx(mask, cv2.MORPH_CLOSE, np.ones((5, 5), np.uint8))
+    contours, _ = cv2.findContours(mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
+    candidates = []
+    area = gray.size
+    for contour in contours:
+        contour_area = cv2.contourArea(contour)
+        if not 0.015 * area < contour_area < 0.985 * area:
+            continue
+        perimeter = cv2.arcLength(contour, True)
+        polygon = cv2.approxPolyDP(contour, 0.018 * perimeter, True)
+        if len(polygon) != 4 or not cv2.isContourConvex(polygon):
+            continue
+        # Use the quadrilateral only to recognize a sheet, then keep its full
+        # contour hull so rounded or slightly uneven edges cannot lose content.
+        polygon = cv2.convexHull(contour)
+        paper = np.zeros_like(gray)
+        cv2.fillPoly(paper, [polygon], 255)
+        inside = cv2.erode(paper, np.ones((7, 7), np.uint8)) > 0
+        ring = (cv2.dilate(paper, np.ones((9, 9), np.uint8)) > 0) & (paper == 0)
+        if inside.sum() < 100 or ring.sum() < 50:
+            continue
+        paper_light = float(np.median(gray[inside]))
+        # A paper edge must separate a mostly light sheet from darker scanner
+        # backing all around it. Ambiguous photos/pages are left untouched.
+        if paper_light < background + 30 or np.mean(gray[ring] < paper_light - 20) < 0.85:
+            continue
+        candidates.append(polygon)
+
+    if not candidates:
+        return image
+    keep_small = np.zeros_like(gray)
+    cv2.fillPoly(keep_small, candidates, 255)
+    # Do not erase another significant object that we could not classify.
+    outside = (keep_small == 0)
+    if np.mean(gray[outside] > background + 30) > 0.04:
+        return image
+    keep = np.zeros((height, width), dtype=np.uint8)
+    polygons = [np.rint(polygon / scale).astype(np.int32) for polygon in candidates]
+    cv2.fillPoly(keep, polygons, 255)
+    # Rounding/downsampling uncertainty grows with the original scan DPI.
+    padding = max(2, int(np.ceil(2 / scale)))
+    keep = cv2.dilate(keep, np.ones((2 * padding + 1, 2 * padding + 1), np.uint8))
+    result = image.copy()
+    result[keep == 0] = 255
+    return result
+
+
+def clean_scan_background(input_path: Path, output_path: Path):
+    image = cv2.imread(str(input_path), cv2.IMREAD_COLOR)
+    if image is None:
+        raise RuntimeError("Scan-Bild konnte nicht gelesen werden")
+    if not cv2.imwrite(str(output_path), whiten_background(image)):
+        raise RuntimeError("Bereinigtes Scan-Bild konnte nicht gespeichert werden")