A fast drop-in alternative to the excellent deskew library by sbrunner, using OpenCV instead of scikit-image for the heavy lifting. It estimates the skew angle of a document image so you can rotate it back to straight.

On sbrunner's own test set it returns the same angles while running ~9× faster (see Benchmark).

• Using pip:

pip install fast-deskew-cv

• From source:

pip install git+https://git@github.com/HOZHENWAI/fast_deskew.git

import cv2
from deskew import determine_skew

image_grayscale = cv2.imread('input_image.png', cv2.IMREAD_GRAYSCALE)
skew_angle = determine_skew(image_grayscale)

determine_skew returns the skew angle in degrees (or None if no dominant orientation is found). To straighten the image, rotate it by +skew_angle:

import cv2
from deskew import determine_skew

image = cv2.imread('input_image.png')
grayscale = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)

angle = determine_skew(grayscale)
if angle is not None:
    height, width = image.shape[:2]
    matrix = cv2.getRotationMatrix2D((width / 2, height / 2), angle, 1.0)
    deskewed = cv2.warpAffine(image, matrix, (width, height),
                              borderValue=(255, 255, 255))
    cv2.imwrite('output_image.png', deskewed)

The classical Hough-transform skew-detection pipeline:

1. Blur the image (Gaussian) to suppress noise.
2. Canny edge detection — thresholds are derived from a percentile of the image's gradient magnitude, so the same defaults work across document sizes and contrasts.
3. Hough line transform to find straight edges.
4. Fold every line orientation into a canonical interval (so text baselines and vertical strokes, which differ by ~90°, reinforce the same estimate) and return the most frequent (modal) angle from the strongest lines.

fast-deskew vs the reference deskew (1.6.1) on the reference library's own 8 test images. Both libraries receive the same grayscale array; timing is the median of repeated calls and excludes image loading. Angles are compared against the ground-truth values from the reference test suite.

Accuracy: 8/8 angles identical to the reference. Speed: ~9× faster (median).

Measured on Python 3.14, OpenCV 4.13, scikit-image 0.26, AMD64. Absolute timings are hardware-dependent; the relative speed-up is the meaningful figure.

• Input is expected to be an 8-bit grayscale image (convert color images with cv2.cvtColor(img, cv2.COLOR_BGR2GRAY) first).
• For skews beyond ±45°/±90°, fix the global page rotation first and use this library for the residual fine skew. angle_pm_90=True widens the fold range.

pip install -e ".[dev]"
pytest

MIT — see LICENSE.

Algorithm and test images from sbrunner/deskew.