Online recognition uses stroke order and writing dynamics; offline recognition depends only on the final image

What it is

Online recognition consumes time-stamped ink: point order, speed, lifts. Offline recognition sees only the bitmap or vector after writing, as if reading a scan. For the same characters, online knows “horizontal then vertical”; offline sees only a crossing. That is a difference in the recognizer’s input channel, not where the correction entry sits, and not editing committed text as keyboard text.

Why it happens

Online features include stroke count, direction sequences, curvature, speed peaks (pauses), and the structure that lifts cut into the path. Many look-alike characters are almost identical as static images and split on stroke order. Offline must render ink to an image, then segment and classify, falling back on appearance and a language model once time is gone. Throwing the online path away and sending only an image voluntarily drops the channel. Conversely, a photograph or a printout has no online signal and must go offline. Hybrid systems often render the path to an image as well and fuse two scores, to survive non-standard stroke order—some people do not follow the prescribed sequence. Dynamics are especially valuable on messy writing: speed and pressure contours remain even when the outline is a blob. Apps that drop samples or store only a final bitmap cannot restore the online channel later.

Studying it

Keep both the path and a rendering of the same ink, and run online, offline, and fused models. Add strokes in deliberately wrong order.

Independent variables: input channel (path / image / both), whether speed is kept, whether stroke order is canonical. Dependent measures: character error rate, drop on shuffled order, inference time.

Do not use scanned print as an online control; that is another kind of data.

Where it stops holding

Photographed whiteboards and homework sheets are offline only. For writers with very non-standard order, online sequence features can hurt, and fusion or appearance-only may win. If legal records keep only an image, an online engine cannot re-recognise later. Encrypted ink that strips timestamps also shuts the channel.

Applying it

• When pen events are available, run online or fused recognition; do not burn ink to an image first.
• Persist the raw path (compressed if needed), not only the rendering, so a new engine can re-recognise.
• Route image-only imports through an offline model, and lower UI expectations of instant characters.
• How to check: CER on the same passage as path versus image-only; canonical order should favour the path. On shuffled samples, see whether fusion holds. Strip timestamps and re-run; results should approach offline.

Related

• Same group: C5.14.2 Segmentation errors cascade into later characters · C5.14.3 Recognition latency trades against accuracy · C5.14.4 Cursive is much less accurate than regular script
• Adjacent: C5.05 Handwriting Recognition and Correction · C5.15 Handwriting Correction and Post-editing
• Search: online handwriting · offline handwriting · stroke order