Identity OCR — French/Arabic Document Intelligence

An observable document pipeline

Identity OCR is Digital Place’s Document Intelligence demonstrator. It explores how a document image becomes structured fields through an interface that exposes processing stages, rather than only returning text. This is an R&D asset, not a customer reference or a KYC decision service.

The public demonstrator presents four Moroccan document families: national identity cards, vehicle registration certificates, driving licences and passports. The tester offers these categories and a beta-labelled automatic detection option. Older and newer variants mentioned in its documentation require qualification on a representative dataset before business integration.

From images to fields

Internal sources describe detection, perspective correction, French/Arabic OCR, MRZ reading where present, quality checks and an operator console. The public interface exposes region-based reading and field extraction. Dual scripts matter: Arabic and Latin names and place names need to be preserved without treating transliteration, correction and proof of identity as equivalent.

Structured output can help prepare a case and reduce re-entry. It should retain links to source regions and support explicit corrections. Format, date and MRZ checksum checks can identify inconsistencies; none establishes document authenticity on its own.

Technical scope and data handling

Internal sources document YOLO, PaddleOCR, OpenCV, FastAPI and React, and describe an on-premise architecture with an operator console. Public documentation states that the engine can run on x86 CPUs without requiring a GPU or an external cloud OCR service. These are documented architectural and deployment characteristics, not an audit of the public server or an independent performance measurement.

Deploying within an information system allows access, networking, logs and image retention to be controlled. Those controls must be implemented and checked for each installation. The online tester receives a file on infrastructure separate from your device: use synthetic or appropriately authorised specimens and review the service’s information before uploading.

Limits of this Lab

The Lab concerns document reading and extraction. It does not claim biometric matching, liveness detection, document fraud detection or automatic KYC eligibility decisions. Identity verification and business decisions require additional controls.

This page publishes no accuracy rate, guaranteed processing time, certification or regulatory compliance claim. Demonstration animations and displayed scores do not replace evaluation against ground truth.

Preparing a useful evaluation

Build an authorised dataset covering document families, variants, scripts and capture conditions. Measure field accuracy, reading failures, human correction rates and latency on target hardware. Test glare, incomplete crops and French/Arabic inconsistencies, then verify that doubtful cases reach human review.

Integration should define the output contract, field provenance, validation rules, access and deletion. This turns a demonstration into a repeatable experiment whose limits can be assessed before it handles a real case.