Image Skew Correction
Two degrees off is enough to hurt — measuring and undoing the tilt in captured pages.
Image skew correction is the estimation and reversal of rotational tilt in document images — the same operation the field also calls document deskewing, approached here from the image-processing side: given a page captured a few degrees off-axis, measure the angle and rotate it back so text baselines run horizontal. Small skews are ubiquitous (paper feeds crooked, cameras tilt, originals were photocopied askew decades ago) and quietly corrosive: line segmentation assumes near-horizontal bands, table cells stop aligning with their grid, and recognition models see character shapes rotated away from their training distribution.
Angle estimation exploits documents' geometric regularity. Projection profiles score candidate angles by how sharply horizontal pixel histograms separate text lines from gaps — the true angle maximizes the contrast. Hough transforms find dominant line orientations from baselines, rules, and table borders. Frequency-domain analysis reads the text-line periodicity's orientation from the Fourier spectrum. Connected-component methods fit baselines through character clusters. Each has failure modes — sparse pages starve projection profiles, decorative diagonals mislead Hough voting — so robust implementations combine estimators or fall back to learned angle regressors on difficult layouts. Correction itself is a rotation with appropriate interpolation and canvas handling; the subtlety is knowing when not to correct — genuinely rotated content (landscape tables in portrait documents, angled stamps) and multi-orientation pages need region-level decisions, not a global spin.
Its adjacent siblings complete the geometric toolkit: orientation detection handles the coarse 90/180/270-degree cases first (skew correction assumes approximate uprightness), perspective correction handles planar tilt from off-axis photography, and dewarping handles the nonlinear curves a single angle can't express. Skew telemetry per intake channel earns its keep operationally — a drift in a branch scanner's skew distribution is a hardware maintenance ticket announcing itself before accuracy visibly drops.
The page went through the scanner at two degrees — deskewing puts the text back on the horizontal.
The photo was taken at an angle — mapping the tilted rectangle back to a flat page.
Before the model reads, the image gets ready — the corrections that decide what recognition has to work with.
Proof Perimeter runs document AI inside your own perimeter — with a provenance record on every field.
Book a demo