
Corrected documents arrive from Processing. Recognition proposes a type with three measured signals — the operator resolves it.
Recognition is a weighted reading of several signals — document shape, aspect ratio, image properties, edges, OCR when needed and local computer vision models. OCR alone is never the deciding input.
The outer contour and corner geometry are read first: a card, a booklet and a sheet present different silhouettes before any pixel is interpreted.
Width-to-height is compared against the stored measurements of each document type. 1.586 is the reference ratio shared by card formats.
Tone distribution, laminate sheen, text density and colour spread separate a photographed card from a scanned page.
Edge strength and continuity locate the document boundary and confirm it survived deskew and perspective correction.
Optical character reading is consulted only when the geometric signals are inconclusive, never as the only input.
Local computer vision models weigh the signals together and return a type candidate with its individual scores.
Recognition does not claim 100% accuracy. Where the signals disagree the type is left unresolved so it can be confirmed or selected manually.
Any AI or computer vision model runs locally on this machine. No cloud service is required and no document image leaves the device.
New document types can be added to this list without rebuilding the program, so operators can register documents as they arrive at the counter.

Recognition runs locally. Images never leave this machine.
Every document passes through the same measured sequence, from the scanner or phone image to the printed sheet or exported file.
DOCUMENT TYPES · TRUE MEASUREMENTS
Every recognised type carries a stored physical size in the measurements database. Recognition proposes the type; the measurement database fixes the millimetres, and the layout never scales a document to fit the sheet.
Recognition reads several signals — shape, aspect ratio, edges, image properties, OCR when it is needed, and a local CV model. All of it runs locally and offline. New document types can be added without rebuilding the program.
The program is offline-first. Recognition, measurement and layout happen locally, and the sheet never rescales a document to make it fit.
Document and boundary detection, perspective correction, cropping, measurement determination, image processing, Auto Layout and print preparation all run on this machine.
Recognition and core processing run locally without an internet connection, so the bench works in an office with no network.
Document images are never sent to external servers; nothing leaves the workstation.
Any AI or computer-vision model used is a local model. No cloud dependency is a condition for the program to work.
Documents are never shrunk or enlarged to fit the sheet; Auto Layout may never silently scale a document.
No comments yet. Be the first!