OCR Processing

Overview

The OCR (Optical Character Recognition) service extracts text from images and PDF documents. It's an independent service that works standalone or combined with other adapters.

Purpose

• Extract text from document images and PDFs

• Support multiple file formats (JPEG, PNG, PDF)

• Provide confidence scores for extraction quality

• Support multiple OCR providers (AWS, Google, Tesseract)

Use Cases

Document Upload & Verification

Upload → OCR Processing → OCR Mapping → Storage → Database

User Registration

Registration → OCR → Mapping → Validation → User Creation

Quick Text Extraction

Upload → OCR Processing → Return Text

Supported Providers

[!NOTE] Test Results: Tested with 4 marksheet documents (English & Hindi). AWS Textract: 95-99% for English, 60-68% for Hindi (garbled). Google Gemini: 90% for both. Tesseract: 25% failure rate. See OCR Provider Comparison Report for full details.

1. AWS Textract

Status: ✅ Production Ready (English documents)

Strengths:

• Excellent English accuracy (95-99%)

• Fast processing (~2 seconds)

• Handles tables and forms well

Limitations:

• Poor Hindi/regional language support (60-68% confidence, garbled output)

• Higher cost

Best For: English-only production workloads

---

2. Google Gemini

Status: ✅ Production Ready (Recommended)

Strengths:

• Excellent multilingual support (English & Hindi: 90%)

• Handles Devanagari script perfectly

• AI-powered context understanding

• Consistent accuracy across languages

Limitations:

• Slower than Textract (~5-6 seconds)

Best For: Hindi/regional languages, mixed language documents, production systems

---

3. Tesseract

Status: ⚠️ Development/Testing Only

Strengths:

• Free and open-source

• Runs locally (no API costs)

• Offline processing

Limitations:

• 25% failure rate in testing

• Poor Hindi support (complete failure)

• Unreliable with complex layouts

Best For: Development/testing, not recommended for production

---

Configuration

Environment Variables

# Choose provider
OCR_PROVIDER=aws-textract    # For English documents
OCR_PROVIDER=google-gemini   # For Hindi/multilingual (recommended)
OCR_PROVIDER=tesseract       # For dev/testing only

AWS Textract

OCR_PROVIDER=aws-textract
AWS_TEXTRACT_AWS_REGION=us-east-1
AWS_TEXTRACT_ACCESS_KEY_ID=your-key
AWS_TEXTRACT_SECRET_ACCESS_KEY=your-secret

Google Gemini

OCR_PROVIDER=google-gemini
GEMINI_API_KEY=your-api-key

Tesseract

OCR_PROVIDER=tesseract
# No additional config needed

Supported File Types

File Type	AWS Textract	Google Gemini	Tesseract
JPEG	✅	✅	✅
PNG	✅	✅	✅
PDF	✅	✅	❌
WebP	❌	✅	❌

Output Format

{
  "fullText": "CERTIFICATE\nName: John Doe\nID: 123456...",
  "confidence": 95,
  "metadata": {
    "pageCount": 1,
    "language": "en",
    "processingTime": 2500
  }
}

Provider Selection Guide

Use AWS Textract When:

• All documents are in English only

• Need fastest processing (4-5 seconds)

• Processing CBSE, ICSE, English medium certificates

• Already using AWS infrastructure

⚠️ Do not use for Hindi/regional language documents

Use Google Gemini When:

• Documents contain Hindi or regional languages

• Mixed language content (Hindi + English)

• Document language is unknown

• Need reliable accuracy across all document types

• Production systems (recommended default)

Use Tesseract When:

• Development and testing environments only

• Budget constraints (no API costs)

• Simple prototyping

⚠️ Not recommended for production

Performance

Actual Test Results (4 Documents)

Metric	AWS Textract	Google Gemini	Tesseract
OCR Time	2.1s	5.8s	2.5s
Total Time	4.7s	7.4s	3.7s
Success Rate	100%	100%	75%
English Accuracy	95-99%	90%	Unreliable
Hindi Accuracy	60-68% (garbled)	90%	Failed

Best Practices

1. Language-Based Routing: Detect language and use appropriate provider

2. Validate File Types: Check file type before processing

3. Handle Errors Gracefully: Return user-friendly error messages

4. Log Performance: Track provider and processing times

5. Implement Retry Logic: Retry failed extractions with exponential backoff

Quality Tips

Improve Accuracy:

• Use high-resolution images (300+ DPI)

• Ensure good contrast and lighting

• Upload straight, non-skewed documents

• Use clean backgrounds

Choose Right Provider:

• English documents → AWS Textract

• Hindi/Unknown → Google Gemini

• Production → Google Gemini (safest choice)

Troubleshooting

Low Confidence / Garbled Text

Cause: Using AWS Textract on Hindi documents

Solution: Switch to Google Gemini

OCR_PROVIDER=google-gemini

Slow Processing

Solution: For English-only, use AWS Textract for faster processing

Provider Not Supported Error

Solution: Check exact provider name:

• aws-textract (not textract)

• google-gemini (not gemini)

• tesseract (not tesseract-ocr)

Security

1. Validate file types to prevent abuse

2. Limit file size (e.g., 10MB max)

3. Sanitize extracted text before storing

4. Store credentials in environment variables only

Summary

The OCR service extracts text from documents using configurable providers:

• AWS Textract: Best for English documents (fast, accurate)

• Google Gemini: Best for Hindi/multilingual (reliable, recommended)

• Tesseract: Development/testing only (not production-ready)

Recommendation: Use Google Gemini as default for production systems to handle all document types reliably.

---

Related Documentation

• Service Adapters - Adapter pattern architecture

• OCR Mapping - AI-powered field extraction

• File Storage - Cloud storage integration

• OCR Provider Comparison Report - Detailed test results and analysis