Training-free page orientation

Deskew before you classify

Level the page first, then let any frozen orientation model vote over its four quarter turns.

Anonymous authors · paper under review

Code

Classify-first versus LevelHead on a page turned 152 degrees — Classified first, a tilted page stays upside down and OCR reads gibberish. Levelled first, it comes out upright, for every backbone (right).

TL;DR Pipelines pick a page's quarter turn first and remove its tilt second, so the classifier sees pages unlike any it was trained on. LevelHead swaps the order and votes over the four quarter turns of the level page. Nothing is trained.

Method

Level the page, ask the backbone four times, roll back and combine — Any angle splits as θ = 90°·k + s. A spectral deskewer removes s; the backbone sees the level page in four turns; answers are rolled back and multiplied, so a fixed preference for one answer cancels out.

Results

Setting	Classify-first	LevelHead	Gain
4 classifiers, pages at any angle	71.9	85.2	+13.3
4 Qwen3-VL models, 7 benchmarks	41.9	84.6	+42.7
GPT-4o-mini, 7 benchmarks	34.3	69.9	+35.6
Gemma-4-31B, 7 benchmarks	62.0	94.5	+32.5
DeepSeek-V4.1-Flash, 7 benchmarks	68.8	96.5	+27.7
4 classifiers, phone photos at any angle	69.3	89.8	+20.5
4 classifiers, DISE scans at any angle	65.0	91.3	+26.3

One page per difficulty level, where the two orders disagree. Same backbone (Qwen3-VL-32B), same vote; only the order differs.

A deskewer that reads photographs

An ink edge at the padding outvotes the text; a Hann taper restores it — The released jdeskew pads a page with white and then thresholds it, so a photo gains an ink line where it meets the padding. That line wins and the answer is 0°. A Hann taper fades it: photos left more than 15° off drop from 31% to 1.3% (screen) and from 33% to 2.6% (oblique). See it live.

Benchmarks

Seven labelled sets

ORB receipts, Indic scripts and documents on photos, plus ODB-4 and ODB-any from all 1,651 OmniDocBench pages.

Phone photos

1,433 Real5 photographs turned by known quarter turns and angles.

Scans

2,800 DISE 2021 test scans with known skew, in all four quarter turns.

Screen photographs corrected by both orders — Screen photographs at arbitrary angles: classify-first leaves them a quarter or half turn off; LevelHead turns them upright.

Runs in your browser

Level any page

Drop in a scan or a phone photo. Step 1 of LevelHead reads its tilt from the Fourier spectrum and levels it, right here, in about a tenth of a second.

Input

Upload

drop an image here

Add tilt0°

Turn The deskewer reads tilt modulo 90°; quarter turns are Step 2's job.

Levelled by Step 1

LevelHead deskewer

…

Hann taper Keep ±45° peak

Tilt read–

Rotation applied–

Time–

Inside the deskewer

Rows of text make a bright streak through the spectrum, perpendicular to the lines. The deskewer sums the spectrum along rays and picks the brightest. Switch the taper off to see the padding edge take over.

Step 2: score the four turns

Loading model

A real orientation classifier, PaddleOCR's PP-LCNet (7 MB), runs on your device. It scores each quarter turn of the level page; each bar is the probability that the view is turned that far counterclockwise.

Roll back into the page's frame

View j was turned j quarter turns, so its answer is read j places along. A correct model now agrees column by column.

Combine with a geometric mean

Final answer–

classify-first

Classify the page as given with one view, then deskew.

LevelHead

Level first, vote over four turns, then turn the page upright.

Everything above runs in this page: the deskewer in a Web Worker, the classifier with ONNX Runtime Web. The paper evaluates many backbones; this demo uses PP-LCNet, the most accurate classifier on level pages.

Explore the results

Turn a page

Drag the angle. See how it splits into a quarter turn and a tilt, and how often each order gets it right.

Page and its level version

Backbone

Page as given

After Step 1: level, a quarter turn off

Classify-first

LevelHead

Accuracy against tilt

ODB-any, 1,651 pages. Both orders use the same deskewer and the same four-view vote.

Every backbone

Backbone	Classify-first	LevelHead	Status

Average accuracy (%) over seven public benchmarks. A model near chance (25%) even on level pages cannot be helped by any wrapper.