auto-deskew

Auto Deskew

Straighten a tilted image. Two backends, picked by content type:

• document (default) — magick -deskew 40%. Detects skew via the Radon-style projection profile that ImageMagick's deskew operator implements. Fast, robust on text/scans up to ±15°.
• photo — Python + OpenCV Hough-line transform. Detects the dominant horizontal/vertical line angle (horizons, building edges, table corners) and rotates by its negative. Use for photographs without a clean text baseline.

When to use

• Scanned documents that came in at an angle.
• Phone snapshots of pages, receipts, whiteboards.
• Slightly-tilted architectural / horizon photos (--mode photo).
• Pre-step before OCR — Tesseract accuracy degrades sharply past ±2° skew.

Do not use this skill when:

• The image is intentionally tilted (Dutch angle, artistic composition).
• Skew is severe (>15°) — usually means the document was captured upside down or sideways. Detect orientation first (exiftool -Orientation, or rotate by 90/180 multiples) before deskewing.
• The image has no straight reference lines (clouds, abstract textures) — Hough will return noise; rotation will be near-random.

Inputs

1. Input — file or directory. Required.
2. Mode — document (default) | photo | auto (heuristic: if magick identify -format "%[colorspace]" is Gray or the image has very low saturation → document; else photo).
3. Recursive — --recursive for directory descent.
4. Output — default in-place suffix _deskew. --output-dir <path> for sibling folder. --overwrite to replace.
5. Threshold (mode=document) — -deskew percentage. Default 40% (ImageMagick's recommended value). Higher = more aggressive detection, more false positives.
6. Max angle — default 15°. If detected skew exceeds this, refuse to rotate and flag the file (probably wrong orientation, not skew).
7. Crop — default on. After rotation, crop to the largest inscribed rectangle so there are no transparent / black triangle borders. --no-crop to keep the full rotated canvas.
8. Background — fill colour for the post-rotation triangles when --no-crop. Default white. Use none for transparent (PNG/WebP only).

Procedure

1. Verify ImageMagick 7+. For photo mode, also verify the plugin's uv venv has opencv-python-headless and numpy (managed by install-deps; if missing, point at it).

2. Enumerate inputs (JPEG/JPG/PNG/TIFF/TIF/WebP). Skip and note RAW.

3. document mode:

   magick "$IN" -background "$BG" -deskew ${THRESHOLD}% \
     $( [ "$CROP" = on ] && echo "+repage -fuzz 1% -trim +repage" ) \
     "$OUT"
   

   -deskew applies the rotation. -trim with a small fuzz removes the resulting fill border when CROP=on. Read back the applied angle from the verbose output (-verbose) for the summary.

4. photo mode (Python via the plugin's uv venv):

   import cv2, numpy as np, sys
   img = cv2.imread(sys.argv[1])
   gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
   edges = cv2.Canny(gray, 50, 150, apertureSize=3)
   lines = cv2.HoughLines(edges, 1, np.pi/720, 200)
   if lines is None:
       print("0.0"); sys.exit(0)
   # Collect angles near horizontal (±30°) and near vertical (±30° of 90°)
   angles = []
   for rho, theta in lines[:200, 0]:
       deg = np.degrees(theta) - 90  # 0 = horizontal line
       if abs(deg) < 30:
           angles.append(deg)
       elif abs(deg - 90) < 30:
           angles.append(deg - 90)
       elif abs(deg + 90) < 30:
           angles.append(deg + 90)
   if not angles:
       print("0.0"); sys.exit(0)
   print(f"{np.median(angles):.3f}")
   

   Then rotate via ImageMagick using the detected angle:

   ANGLE=$(uv run --project "$PLUGIN_VENV" python deskew_photo.py "$IN")
   magick "$IN" -background "$BG" -rotate "$(awk "BEGIN{print -1*$ANGLE}")" \
     $( [ "$CROP" = on ] && echo "+repage -fuzz 1% -trim +repage" ) \
     "$OUT"
   

5. Max-angle guard: if |angle| > MAX_ANGLE, skip the file and flag it: "$IN: detected ${angle}° — exceeds max-angle ${MAX_ANGLE}°, skipping (likely wrong orientation, not skew)".

6. Crop to inscribed rectangle (when --no-crop is off and the simple -trim leaves uneven borders): use the analytic formula for the largest inscribed axis-aligned rectangle inside a rotated rectangle. For input W×H rotated by θ:

   new_w = (W·|cos θ| - H·|sin θ|) / (cos²θ - sin²θ)   if W ≥ H
   new_h = (H·|cos θ| - W·|sin θ|) / (cos²θ - sin²θ)
   

   Crop centred. For small angles (<5°) the difference vs. -trim is negligible — -trim is fine.

7. For batch, parallelise with xargs -P $(nproc). Skip files ending _deskew.

Output

• Deskewed images at the resolved paths.
• Summary:
  • Files processed / skipped (RAW or non-image) / failed / unchanged (|angle| < 0.1°) / refused (|angle| > max).
  • Per-mode count when mixed.
  • Distribution of detected angles (min / median / max) — sanity check across the batch.

Notes

• ImageMagick's -deskew is tuned for bilevel/grayscale text. On colour photos it often returns 0° or noise — that's why photo mode exists.
• For OCR pipelines: deskew before binarisation. After deskew, magick "$IN" -threshold 50% or hand off to tesseract directly.
• If you want both content types handled in one run: --mode auto. The heuristic isn't perfect — review the angle distribution in the summary and re-run misclassified files explicitly.
• This skill rotates only. For perspective correction (four-corner unwarping of a photographed document), use a dedicated tool like dewarp or OpenCV's getPerspectiveTransform — auto-deskew won't fix keystoning.