OCR Engine

Turqoa's OCR engine is a purpose-built optical character recognition system optimized for port and terminal environments. It handles three distinct recognition tasks — license plates, container codes, and seal numbers — each with specialized models trained on millions of real-world port images.

Capabilities

License Plate Recognition

The plate OCR model recognizes license plates from over 120 countries and regional formats. It handles:

• Standard passenger and commercial vehicle plates
• Multi-line plate formats (common in European and Middle Eastern regions)
• Temporary and dealer plates
• Dirty, damaged, or partially obscured plates
• Plates captured under infrared illumination

Container Code Recognition

The container code model is trained specifically on ISO 6346 container markings. It extracts:

• Owner code — 3 letters identifying the container operator
• Equipment category — 1 letter (U, J, or Z)
• Serial number — 6 digits
• Check digit — 1 digit, validated automatically against ISO 6346

The model also reads:

• Container size and type codes
• Maximum gross weight markings
• Tare weight markings
• CSC plate data (when visible)

Seal Recognition

The seal OCR model reads alphanumeric codes from high-security bolt seals, cable seals, and e-seals:

• ISO 17712 compliant seal formats
• Carrier-specific seal numbering
• QR and barcode seals (via separate barcode decoder)

Supported Formats

Container Codes (ISO 6346)

Note: The engine automatically validates the check digit using the ISO 6346 algorithm. Mismatched check digits are flagged and the transaction confidence is reduced.

Regional Plate Formats

Turqoa ships with recognition profiles for all major port regions:

Configuration

# ~/.turqoa/sites/my-terminal/ocr.yaml
ocr:
  container:
    model: turqoa-container-v4
    confidence_threshold: 0.90
    check_digit_validation: true
    retry_on_low_confidence: true
    max_retries: 2

  plate:
    model: turqoa-plate-v3
    confidence_threshold: 0.85
    regions:
      - middle_east
      - europe
    multi_line: true
    infrared_mode: auto

  seal:
    model: turqoa-seal-v2
    confidence_threshold: 0.80
    barcode_fallback: true
    formats:
      - iso_17712
      - carrier_specific

Confidence Thresholds

Confidence thresholds control when OCR results are considered reliable enough for automated processing. Setting these correctly is critical for balancing automation rate against accuracy.

Tune thresholds using historical data:

turqoa ocr analyze --site "my-terminal" --period 7d

This produces a report showing accuracy at different threshold levels, allowing data-driven tuning.

Multi-Language Support

The OCR engine natively supports multiple scripts:

• Latin — English, Spanish, Portuguese, French, German, Turkish
• Arabic — Standard Arabic, Farsi numerals
• CJK — Simplified Chinese, Traditional Chinese, Japanese, Korean
• Cyrillic — Russian, Ukrainian
• Devanagari — Hindi

For plates with dual-script formats (common in the Middle East), the engine reads both scripts and cross-validates:

ocr:
  plate:
    dual_script:
      enabled: true
      primary: arabic
      secondary: latin
      cross_validate: true

Performance Tuning

Image Quality

OCR accuracy is directly tied to image quality. Key factors:

• Resolution — Minimum 150 pixels across the plate or container code
• Exposure — Avoid overexposed or underexposed captures
• Focus — The character region must be in sharp focus
• Angle — Maximum 30-degree skew from perpendicular

Processing Speed

To optimize throughput on multi-lane installations:

ocr:
  processing:
    batch_size: 4          # Process multiple images per GPU call
    worker_threads: 2      # Parallel processing threads per model
    gpu_memory_fraction: 0.8  # Reserve 80% of GPU for OCR
    preload_models: true   # Load models at startup, not on first request

Warning: Setting gpu_memory_fraction too high may starve the damage detection model of GPU resources. Ensure total GPU memory allocation across all models does not exceed 90% of available VRAM.

Retry Logic

When a read falls below the confidence threshold, the engine can request a re-capture:

ocr:
  retry:
    enabled: true
    max_attempts: 3
    delay_ms: 500
    escalate_on_failure: true  # Route to manual review after all retries fail