Deskewing

Deskewing is the process of correcting skew in an image or document so that features intended to be horizontal or vertical align with the axes. It is a common preprocessing step in digital document handling, scanned archives, and computer vision, with the goal of improving readability and subsequent analysis such as optical character recognition (OCR).

In document imaging, deskewing detects the skew angle of text lines and rotates the image to compensate. The angle is typically small, measured in degrees from -5° to 5° but can be larger for severely misaligned scans. Methods include projection-profile analysis, which seeks the angle that minimizes vertical projections; Hough or Radon transforms that detect linear alignments; Fourier-based methods and image moment techniques. Some approaches combine multiple cues or use learning-based predictors.

Once the angle is determined, the image is rotated by the negative of that angle using an interpolation scheme (nearest neighbor, bilinear, bicubic). Rotation may introduce artifacts or blur and can change the image size, often requiring cropping to remove introduced borders.

Deskewing can be automatic in scanning software and document editors, or performed as a preprocessing step in OCR pipelines. It may also be used in other imaging contexts where alignment matters, such as aerial or medical images, though the specific techniques differ.

Advances include deep learning models that estimate rotation or perform end-to-end deskewing without explicit angle estimation. Evaluations focus on accuracy of angle estimation and preservation of text quality.