Skip to content

Document Capture & Quality

Definition

Document capture is the process of acquiring a high-quality image of an identity document through a phone camera or scanner. Quality at capture directly determines the accuracy of all downstream processing — OCR, forensics, and face extraction.

Auto-Capture System

graph TD
    A[Camera Preview] --> B[Real-Time Analysis]
    B --> C{Document detected?}
    C -->|No| D["Guide: 'Place document in frame'"]
    C -->|Yes| E[Quality Checks]

    E --> F{All checks pass?}
    F -->|No| G["Guide: 'Move closer' / 'Reduce glare' / 'Hold steady'"]
    F -->|Yes| H[Auto-Capture Triggered]
    H --> I[High-res image saved]

    D --> A
    G --> A

    style H fill:#2E7D32,color:#fff

Quality Checks

Check Metric Threshold Detection Method
Document presence Detection confidence > 0.9 Object detection model
Full document visible All 4 corners detected All inside frame Corner detection
Blur Laplacian variance > 100 Laplacian filter
Glare/reflection Specular highlight area < 5% of doc area Highlight detection
Shadow Luminance uniformity Variance < threshold Regional brightness analysis
Resolution DPI equivalent > 300 DPI Pixel density calculation
Tilt/skew Angle from horizontal < 10° Corner-based angle computation
Distance Document size in frame 60-90% of frame Bounding box ratio
Lighting Mean brightness 80-220 Histogram analysis
Occlusion Finger/thumb overlap < 3% overlap Hand detection model

Perspective Correction

After capture, apply perspective transform to get a flat, rectangular document image:

import cv2
import numpy as np

def perspective_correct(image, corners):
    """
    corners: 4 detected document corners [TL, TR, BR, BL]
    Returns: perspective-corrected rectangular image
    """
    # Target dimensions (standard ID card aspect ratio)
    width, height = 856, 540  # ~CR80 card at 100 DPI

    dst = np.array([[0, 0], [width, 0], [width, height], [0, height]], dtype=np.float32)
    M = cv2.getPerspectiveTransform(corners.astype(np.float32), dst)
    corrected = cv2.warpPerspective(image, M, (width, height))
    return corrected

Front vs Back Capture

Side What's Extracted Special Challenges
Front Photo, name, DOB, document number, expiry Hologram glare over photo area
Back MRZ, barcode, additional data, signature Low contrast text, barcode readability

Key Takeaways

Summary

  • Quality at capture determines everything — poor capture = poor OCR + poor forensics
  • Auto-capture with real-time guidance dramatically improves first-attempt success rate
  • Critical checks: blur, glare, all corners visible, resolution > 300 DPI
  • Perspective correction after capture normalizes the document for downstream processing
  • Both front and back capture are typically needed for full data extraction