Document Verification Deep Dive

Document capture is more important than most teams realize. Poor image quality is the leading cause of verification failures and manual review escalations.

Common capture problems:

• Blur from camera shake or poor focus
• Glare from reflective ID surfaces
• Shadows obscuring text or photos
• Cropped corners missing security features
• Low resolution making text unreadable

Guided capture best practices:

• Real-time feedback on image quality
• Auto-capture when quality thresholds met
• Clear instructions with visual guides
• Multiple capture modes (camera, upload)
• Fallback to manual capture with review

Investing in capture quality pays dividends throughout the verification pipeline. A clear image makes every subsequent step more accurate.

Before extracting data, the system must identify what type of document it's looking at.

Classification challenges:

• 195+ countries with varying ID formats
• Multiple document types per country
• Regional variations and updates over time
• Front vs back detection
• Two-sided vs single-sided documents

Classification approaches:

• Template matching: Compare against known document templates
• ML classification: Train models on document features
• Hybrid: Use templates for common documents, ML for edge cases

Output of classification:

• Document type (passport, national ID, driver's license)
• Country of issuance
• Document version/template
• Which fields to extract and validate

OCR (Optical Character Recognition) extracts text from document images. Identity documents present unique challenges.

Standard field extraction:

• Full name (first, middle, last)
• Date of birth
• Document number
• Expiration date
• Address (where present)

MRZ (Machine Readable Zone): Passports and some IDs include MRZ—standardized text blocks with check digits. MRZ parsing provides:

• Structured data extraction
• Built-in error detection via check digits
• High accuracy even on lower quality images

Challenges and solutions:

• Fonts: ID-specific fonts differ from standard OCR training data
• Languages: Non-Latin scripts, diacritics, transliteration
• Layout: Field positions vary by document type
• Damage: Worn, scratched, or faded text

Purpose-built document OCR significantly outperforms general-purpose OCR on identity documents.

Verifying that a document is genuine—not forged, altered, or fraudulently obtained.

Physical security features (detected visually):

• Holograms and optically variable devices
• Microprinting (tiny text visible under magnification)
• Security patterns and guilloche
• UV-reactive elements
• Raised lettering/embossing

Digital analysis:

• Compression artifact analysis
• Font consistency checking
• Photo manipulation detection
• Template structure validation
• Color profile analysis

Common fraud patterns:

• Digital manipulation: Photoshopped text, swapped photos
• Physical forgery: Fake documents printed on standard paper
• Stolen blanks: Genuine blanks obtained illegally
• Compromised documents: Real documents with fraudulent data

Modern systems combine multiple detection methods. No single check catches all fraud—defense in depth is essential.

Beyond document authenticity, validate that the document and data are legitimate.

Expiration checking:

• Document hasn't expired
• Issue date is plausible
• Document age appropriate for holder's birth date

Database verification:

• Government database lookups (where available)
• Lost/stolen document registries
• Sanctions and watchlist screening

Data consistency:

• Extracted data matches user-provided data
• Internal document consistency (e.g., MRZ matches visual zone)
• Cross-document consistency for returning users

Third-party data:

• Address verification services
• Phone number validation
• Email reputation checking

The goal is building confidence across multiple independent signals, not relying on any single verification.