Skip to main content

Amazon Rekognition: Computer Vision, Custom Labels & Content Moderation

Key Takeaways​

Amazon Rekognition is a fully managed, serverless Computer Vision (CV) service that allows developers to analyze images, stored video files, and live streaming video feeds via pre-trained deep learning APIs without needing machine learning expertise.

Out of the box, Rekognition provides Object/Scene Detection, Facial Analysis & Verification, Face Liveness Detection (anti-spoofing), Text in Image Detection, PPE Detection, and Content Moderation. For specialized use cases, it provides Custom Labels (recognizing proprietary logos and custom objects with small training sets) and Custom Moderation Adapters, with seamless integration into Amazon Augmented AI (Amazon A2I) for human review workflows.

Main Discussion​

Core Pre-Trained Computer Vision Capabilities​

FeaturePrimary MechanismTypical Business Application
Object, Scene & Label DetectionIdentifies thousands of common objects (vehicles, pets, tools) and environmental scenes (outdoors, beach).Auto-tagging media assets, searchable digital asset management (DAM) libraries.
Facial Analysis & ComparisonDetects faces, estimates age ranges, emotions (happy, calm), eye state (open/closed), and compares similarity across photos.User onboarding, photo organization, entry security checkpoints.
Face Liveness DetectionVerifies that a real, physically present user is interacting with the camera, blocking printed photos, screen replays, and 3D mask spoof attacks.Fraud prevention and identity verification in mobile banking and authentication apps.
Facial Search (Collections)Indexes facial vectors into managed Face Collections in Amazon Rekognition to match faces across large databases.Employee access control, VIP guest identification, suspect lookup in authorized security feeds.
Text DetectionExtracts printed and handwritten text across various orientations and fonts in images.Reading highway signs, license plates, race bib numbers, or billboard advertisements.
Personal Protective Equipment (PPE)Identifies face covers, hand covers, and head covers on persons, checking if equipment covers the intended body parts.Industrial safety audits, automated workplace OSHA compliance monitoring.
Celebrity RecognitionRecognizes known actors, athletes, politicians, and public figures.Media indexing, cataloging broadcast television footage, marketing analytics.
Person PathingTracks the spatial movement coordinates and trajectory of individuals across video frames over time.Sports performance tracking, retail foot-traffic pattern analysis, physical security.

Amazon Rekognition Custom Labels​

When standard pre-trained models cannot recognize specialized business products, proprietary logos, or industrial defects, Custom Labels enables transfer learning with minimal training overhead:

  • Low Data Requirement: Requires only a small labeled dataset (often fewer than a few hundred images) to achieve high domain accuracy.
  • No Machine Learning Code: Amazon Rekognition handles model selection, transfer learning, hyperparameter optimization, and endpoint hosting behind the scenes.

Automated Content Moderation & Human-in-the-Loop (Amazon A2I)​

Amazon Rekognition provides granular content moderation to protect brand reputation and user safety across user-generated content (UGC) and generative AI outputs:

  1. Automated Filtering: The DetectModerationLabels API evaluates content across categories (nudity, violence, hate symbols, suggestive content, drugs/tobacco). It handles 95% to 99% of images automatically.
  2. Amazon Augmented AI (Amazon A2I): For ambiguous or low-confidence predictions (typically 1% to 5% of total volume), Rekognition routes the image to Amazon A2I for human verification.
  3. Custom Moderation Adapters: Organizations can train custom adapters using domain-annotated training data to adjust sensitivity thresholds for region-specific compliance rules.

Exam Guide​

Exam Tips​

  • Core Service Purpose: Amazon Rekognition is the primary AWS service for image and video analysis, object detection, facial recognition, and content moderation.
  • Custom Object / Logo Detection: If a scenario asks to detect proprietary brand logos, custom manufactured parts, or specific store products using a small set of sample images, choose Amazon Rekognition Custom Labels.
  • Anti-Spoofing Face Verification: If an exam question asks how to verify that a user in a facial authentication workflow is physically present and not holding up a printed photo or playing a video loop, select Amazon Rekognition Face Liveness Detection.
  • Human Review Integration: The designated AWS service for routing low-confidence machine learning predictions (such as ambiguous moderation flags) to human reviewers is Amazon Augmented AI (Amazon A2I).
  • Face Collections: Rekognition does not store raw photos when building face registries; it extracts and indexes mathematical vector representations into a Face Collection.

Practice Test​

Question 1​

A marketing intelligence platform wants to track how often its corporate logo appears in user-uploaded event photos on social media. The company has 150 sample photos of the logo and wants a managed solution with minimal operational overhead and no custom model scripting. Which AWS solution meets these requirements?

  • A. Train a custom convolutional neural network from scratch on Amazon EC2 GPU instances
  • B. Use Amazon Rekognition Custom Labels trained on the sample logo images
  • C. Use Amazon Textract Key-Value extraction
  • D. Configure an Amazon Comprehend Custom Entity Recognizer
Correct Answer
  • B. Use Amazon Rekognition Custom Labels trained on the sample logo images
    • Explanation: Amazon Rekognition Custom Labels allows developers to train a custom object detection model using a small labeled dataset (in this case, 150 images of the corporate logo) without needing to write machine learning code or manage infrastructure.

Question 2​

A social media platform processes millions of daily photo uploads using the Amazon Rekognition DetectModerationLabels API to remove inappropriate content. For borderline images where the machine learning confidence score is between 60% and 80%, the platform requires human moderators to review the images before final publication. Which AWS service provides this human review workflow integration?

  • A. Amazon QuickSight
  • B. Amazon Augmented AI (Amazon A2I)
  • C. AWS Step Functions with Amazon Polly
  • D. Amazon CloudWatch Synthetics
Correct Answer
  • B. Amazon Augmented AI (Amazon A2I)
    • Explanation: Amazon Augmented AI (Amazon A2I) integrates directly with Amazon Rekognition to route low-confidence or ambiguous machine learning predictions to human review workflows automatically.