Skip to main content
Welcome back to Solutions Architect. Instructor: Michael Forrester. In this lesson we’ll revisit the power of inference using Amazon Rekognition — a fully managed image and video analysis service from AWS. Rekognition provides ready-made computer vision capabilities (label detection, face analysis, OCR, moderation, and more) so you don’t need to build or train deep neural networks to extract actionable metadata from images and videos. It’s widely used in production systems and frequently appears on AWS certification exams. At a high level, Rekognition automatically inspects images and videos and returns structured metadata and suggested tags that describe the content. For example, an image of a dog on a skateboard might be tagged with dog, skateboard, outdoor, and action. This automated tagging is especially valuable for large media libraries where manual labeling is expensive and slow.
A diagram showing an image being processed by Amazon Rekognition. The service outputs suggested keywords like Mountain, Peak, Range, Glacier, Valley, Sunny, Himalayas, Andes, Panorama, and Landscape.
How Rekognition helps
  • Extracts labels, bounding boxes, and attributes from images
  • Detects and reads text (OCR) within images and video frames
  • Analyzes faces (landmarks, estimated age range, emotions, and more)
  • Tracks people and activities in video with timestamps and bounding boxes
  • Flags unsafe or inappropriate content for moderation workflows
Common use cases
  • Automated content organization and tagging for large image/video repositories
  • Search and discovery (indexing images by detected labels and text)
  • Content moderation at scale (automatically flagging potential policy violations)
  • Face verification and search (match faces against stored collections)
  • Real-time analytics using streaming input (Kinesis + Rekognition Streaming)
Content moderation is a frequent and important application. Platforms such as Reddit, Pinterest, and Instagram rely on automated moderation to scale. Rekognition can detect potentially unsafe or inappropriate content and flag images or videos so that automated rules or human reviewers can take action.
A diagram titled "Rekognition" showing a robot analyzing image thumbnails and sorting them into two dotted boxes labeled "Harmful Content" (orange) and "Safe Content" (green).
Integration and typical architecture
  • Event-driven invocation: Use S3 object-created events to trigger processing (via Lambda, Step Functions, or container tasks).
  • Asynchronous video analysis: Rekognition Video supports asynchronous jobs that persist results and provide timestamps for person tracking and activity detection.
  • Results persistence: Store labels, bounding boxes, timestamps, and metadata in DynamoDB, S3, or a search index (Elasticsearch/OpenSearch) for fast querying and retrieval.
  • Human review: When confidence is low, route results to Amazon Augmented AI (A2I) so humans can verify or correct outputs.
  • Streaming: For low-latency, real-time analysis, integrate Kinesis Video Streams with Rekognition Streaming APIs.
A diagram titled "Rekognition" showing an image uploaded to an Input S3 bucket that triggers an AWS Lambda function, which calls Amazon Rekognition and stores the results in Amazon DynamoDB.
Use Rekognition confidence scores to decide automated vs human review paths. For example, send results below a chosen confidence threshold (e.g., < 80%) to A2I for human verification and accept higher-confidence results automatically.
Key features and capabilities Sample Rekognition labels response (simplified)
Best practices
  • Use confidence thresholds to separate automated actions from items needing human review.
  • Persist detection metadata (labels, timestamps, bounding boxes) to a searchable store for fast retrieval and analytics.
  • For specialized detection tasks, train a Custom Labels model instead of relying solely on built-in labels.
  • Monitor costs: video analysis and real-time streaming can increase costs—use sampling, batching, and appropriate retention policies.
  • Apply least-privilege IAM policies for Rekognition and any resources (S3 buckets, Lambda, DynamoDB) interacting with it.
An infographic titled "Features" showing ten numbered cards. Each card lists an image-analysis capability (e.g., Object & Scene Detection, Facial Analysis & Recognition, Text in Image, Emotion Detection, Real-Time Analysis, etc.).
That’s Rekognition — a managed, scalable service for extracting rich metadata from images and videos and for automating moderation, indexing, and search workflows. Links and references

Watch Video