Amazon Rekognition

• Amazon Rekognition
  • Overview
  • Recognition Custom Labels
  • Content Moderation

Overview

• Find objects, people, text, scenes in images and videos using ML
• We can do facial analysis and facial search to do user verification, people counting
• We can create a database of “familiar faces” or compare against celebrities
• Use cases:
  • Image moderation
  • Image labeling
  • Image content moderation
  • Facial analysis
  • Text Detection
  • Celebrity recognition
  • Face detection and analysis (gender, age range, emotions…)
  • Video analysis (object/people/celebrities marked on a timeline, people pathing)
• Input images can come from S3 or provided as bytes as part of the request
• Facial recognition depends on good lightning, angle, visibility of eyes, resolution
• Video must come from Kinesis Video Streams
• Commonly integrated with AWS Lambda: Example: trigger image analysis automatically when an image is uploaded to S3

Recognition Custom Labels

• Allows you to train Rekognition with your own labeled images
• Useful for detecting custom products, logos, or objects
• Workflow:
  • Label training images
  • Store images in Amazon S3
  • Train Rekognition to create a Custom Labels model
  • Analyze new images to detect trained items
• Example: NFL uses custom labels to identify team logos, pylons and foam fingers in images

Content Moderation

• Automatically detects inappropriate, unwanted, or offensive content
• Common use cases:
  • Social media image filtering
  • Broadcast media
  • Advertising verification
  • Child-safe platforms
• Integrates with Amazon Augmented AI (Amazon A2I) for human review
• Custom Moderation Adaptor:
  • Extends Rekognition capabilities by providing our own labeled set of images
  • We can enhance the accuracy of Content Moderation or create a specific use case of moderation
  • Workflow:
    • Label images
    • Train a Custom Moderation Adapter
    • Images pass/fail automatically
    • Uncertain cases → Amazon A2I
    • Feedback improves future moderation
• Example: Content Moderation API
• Scenario: chatbot generates images for users
  • Before returning the image:
    • Call DetectModerationLabels API
    • Rekognition analyzes image safety
    • If safe → return image
    • If unsafe → block or send for human review