AWS Lambda OCR Service

A serverless OCR (Optical Character Recognition) service built with AWS Lambda, API Gateway, Textract, and Rekognition that processes base64 images and returns the text with highest confidence score.

Architecture

API Gateway → Caller Lambda → Target Lambda → AWS Textract + Rekognition

• Caller Lambda (proxy_service): Receives HTTP requests and invokes target Lambda
• Target Lambda (text_recognition_service): Performs dual OCR using AWS Textract and Rekognition
• API Gateway v2: HTTP API endpoint for external access

Features

• Dual OCR Processing: Uses both AWS Textract and Rekognition for maximum accuracy
• Confidence Scoring: Returns result with highest confidence score
• Base64 Image Input: Accepts images as base64 strings via HTTP POST
• Error Handling: Graceful fallback if one OCR service fails
• CloudWatch Logging: Full request/response logging for debugging

Quick Start

Prerequisites

• AWS CLI configured
• Terraform installed
• Node.js 18+ installed

Directory Structure

project/
├── main.tf
├── proxy_service/
│   └── index.js
└── text_recognition_service/
    ├── index.js
    ├── package.json
    └── node_modules/

Lambda to Lambda communication

Comparison Table

Method	Latency	Coupling	Scalability	Cost	Complexity	Use Case
Direct Invoke	Low	High	Medium	Low	Low	Simple request/response
SQS	Medium	Low	High	Low	Medium	Decoupled async processing
SNS	Medium	Low	High	Low	Medium	Fan-out notifications
EventBridge	Medium	Low	High	Medium	Medium	Event-driven architecture
Kinesis	Medium	Low	Very High	Medium	High	Stream processing
DynamoDB Streams	Medium	Low	High	Low	Medium	Data change reactions
S3 Events	Medium	Low	High	Low	Low	File processing
API Gateway	Medium	Medium	High	Medium	Medium	HTTP-based communication
Step Functions	Medium	Low	High	High	High	Complex workflows
Destinations	Low	Medium	Medium	Low	Low	Success/failure routing

Best Practices by Use Case:

• Simple sync communication: Direct invoke (RequestResponse)
• Async processing: SQS or SNS
• Event-driven architecture: EventBridge
• High-throughput streaming: Kinesis
• Complex workflows: Step Functions
• File processing: S3 Events
• Web APIs: API Gateway
• Database triggers: DynamoDB Streams

I have used direct Invoke for simplicity.

Setup

1. Install dependencies:

cd text_recognition_service
npm install
cd ..

2. Deploy infrastructure:

terraform init
terraform apply

3. Get API URL:

terraform output api_url

Usage

API Endpoint

POST https://your-api-url/serverless_lambda_stage/process
Content-Type: application/json

{
  "image": "base64_encoded_image_string"
}

Example Request

curl -X POST https://your-api-url/serverless_lambda_stage/process \
  -H "Content-Type: application/json" \
  -d '{"image": "iVBORw0KGgoAAAANSUhEUgAAAZAAAADI..."}'

Example Response

{
  "message": "Image processed successfully",
  "result": {
    "success": true,
    "data": {
      "bestResult": {
        "source": "textract",
        "text": "Hello World Document",
        "confidence": 0.95,
        "wordCount": 3
      },
      "allResults": [
        {
          "success": true,
          "source": "textract",
          "text": "Hello World Document",
          "confidence": 0.95,
          "wordCount": 3
        },
        {
          "success": true,
          "source": "rekognition",
          "text": "Hello World Document",
          "confidence": 0.87,
          "wordCount": 3
        }
      ],
      "processingInfo": {
        "imageSize": 15420,
        "imageType": "jpeg",
        "timestamp": "2025-01-11T14:30:00.000Z",
        "servicesUsed": ["textract", "rekognition"]
      }
    }
  }
}

Configuration

Environment Variables

• TARGET_LAMBDA_FUNCTION_NAME: Auto-configured by Terraform
• AWS_REGION: Set to eu-central-1

AWS Resources Created

• 2x Lambda functions (caller + target)
• API Gateway v2 (HTTP API)
• S3 bucket for Lambda code storage
• IAM roles and policies
• CloudWatch log groups

Permissions Required

• lambda:InvokeFunction
• textract:DetectDocumentText
• rekognition:DetectText
• logs:*

Monitoring

CloudWatch Logs

• Caller Lambda: /aws/lambda/caller-lambda
• Target Lambda: /aws/lambda/target-lambda
• API Gateway: /aws/api_gw/serverless_lambda_gw

View Logs

# Get log URLs
terraform output caller_lambda_logs
terraform output target_lambda_logs

Cost Optimization

• Textract: ~$1.50 per 1000 pages
• Rekognition: ~$1.00 per 1000 images
• Lambda: Pay per invocation + execution time
• API Gateway: Pay per request

Troubleshooting

Common Issues

1. Internal Server Error

   • Check CloudWatch logs for detailed error messages
   • Verify IAM permissions for Textract/Rekognition

2. Code Not Updating

   • Ensure source_code_hash is set in Terraform
   • Force update: terraform apply -target=aws_lambda_function.target_lambda

3. OCR Not Working

   • Verify image is valid base64
   • Check image format (supports PNG, JPEG, GIF)
   • Ensure image size < 5MB for Textract, < 5MB for Rekognition

Debug Commands

# Check Terraform plan
terraform plan

# View current Lambda code
aws lambda get-function --function-name target-lambda

# Test Lambda directly
aws lambda invoke --function-name target-lambda \
  --payload '{"image":"base64string"}' response.json

Development

Local Testing

Test individual components:

# Test image generation
# Use the provided HTML tool to generate base64 test images

# Test Lambda locally (with AWS SAM)
sam local invoke target-lambda -e test-event.json

Adding New OCR Services

1. Add new OCR function in text_recognition_service/index.js
2. Update Promise.allSettled array
3. Add required IAM permissions
4. Update processOCRResults function

License

MIT License